← Back to leaderboard
haider1

Haider.

Verifiedactive
@haider1

open intelligence for a decentralized future

RosterAI Model News & ReleasesPolitics & Current Events
72,490 followers · 4,310 following · Last polled 3h ago
Perf. Score
66.1
Rank #98
Value Score
52.4
#126 of 273 priced
Median views
7.1K
p25 5.0K · trailing 7d
Followers
72K
+0.6% 7d
ER (impr.)
2.49%
eng ÷ impressions
ER (foll.)
0.54%
eng ÷ followers
Est. CPM
$35.45
QT-rate basis
WoW views
-48.1%
▼ falling

Rates & value · est. from 7.1K median organic views

fairly priced
Quote tweetbasis
$250
$35.45 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 52.4 · rank #126 of 273 priced$1432.66 / 1K engagementsrate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (7.1K), not the mean (16K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (5.0K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
Mathematician Terence Tao: "we're seeing the end of the romantic era of mathematics, the age of heroes. we're moving into an industrial era where math becomes large-scale, done in big teams like the other sciences" We're entering, rather unwillingly, a new era where exciting new things are possible, but some things are being lostfrozen7d ago1.1M4.8K5412531851.8K0.72%
the upcoming chatgpt pricing structure is basically: $20 = the new free tier $100 = the new plus tier ($20) $200 = the new pro 5x tier ($100) $500 = the new pro 20x tier ($200) as simple as thatfrozen28d ago416K2.8K52189272700.81%
this aged terribly. like, very terribly https://t.co/bFirTHVoFgfrozen24d ago359K6.8K15580171862.03%
Fields Medalist Richard Borcherds on OpenAI proofs: "i think what people are getting upset about is that they'e slowly realizing that they are dumb and somehow blaming this on AI, even though it's not really AI's fault" finally, someone who gets it https://t.co/NJmGFTMvyj3d ago142K2.7K27554345932.58%
Mathematician Terence Tao: "one thing this whole era of AI is teaching us is that our idea of intelligence is not really accurate" The history of AI has been: 'here’s a task only humans can do', and then one by one, an algorithm does it too Maybe we're not quite as smart as we thinkfrozen4d ago131K1.8K202133726962.24%
Ex-Google CEO Eric Schmidt: "the belief is among some people in San Francisco that there will be an explosion in learning that will accelerate past what humans can understand" Imagine a mathematical proof that's too complex for any human to understand, yet we can verify it's true What does it mean if humans can never understand it?frozen12d ago130K6138771312880.84%
absolutely pathetic "Terence Tao reposts a statement from the Association for Human Mathematics urging mathematicians to stop working with OpenAI" i keep saying tao has a massive ego (and he's earned it) but he wants a small group to control the entire field. thankfully that's not going to happen anymorefrozen14d ago88K1.5K97120331962.19%
WTF... Gemini 4 Argon ABSOLUTELY murdered opus 5.5, Fable 5 and GPT-6 Astra https://t.co/q3X4oOIanVfrozen22d ago83K64938558490.96%
opus 5.5 is behaving like a very different model today it had been flawless since launch, but now it's dropping tasks, skipping commits, and claiming to have done work it hasn't. even its deep reasoning seems to have turned into basic pattern matching i wouldn't call it a nerf yet, but is anyone else experiencing this?frozen21d ago81K517121109580.87%
gemini 4 argon scores 53 on the Artificial Analysis Intelligence Index tied with gpt-6 astra and fable 5.1, while 1 point ahead of gpt-6.1 sol https://t.co/Yp28HiKYBdfrozen22d ago78K143123280.21%
Yann LeCun says: "despite all the wonderful things AI can do today, we're still very far from matching human and animal intelligence. it's not just around the corner — it's not going to happen in the next 2 years" Your cat understands the physical world better than today's most powerful AI systemsfrozen20d ago77K47567166251301.12%
i'm a software engineer but mathematicians might actually be in a worse spot than us SWE still requires taste, design, intuition and real-world context that AI often struggles with. plus someone needs to understand what they're shipping math on the other hand is more about following logic toward a clear goal which seems easier for AI to handlefrozen11d ago74K1.1K45101111191.84%
woah, the biggest upgrade in gemini 4 argon will be hallucination reduction compared to 3.8 flash, google has basically crushed one of gemini’s biggest weaknesses hallucinations now close to being eliminated https://t.co/SSdKYzC2vifrozen22d ago71K1.2K8927141382.03%
nah sorry, sonnet 5.5 is a disaster in cost here on the AA index sonnet 5.5 max — $7.60/task gpt-6 astra max — $3.26/task sonnet 5 xhigh — $2.87/task so 5.5 max is more than 2x Astra's cost per Intelligence index task https://t.co/RDISXVO2Gzfrozen24d ago67K310186010450.66%
sonnet 5.5 is an absolute token-hungry monster 5.5 max — 193k output tokens/task opus 5.5 max — 119k sonnet 5 xhigh — 65k gpt-6 sol max — 31k gpt-6 astra max — 27k using 62% more output tokens than opus 5.5 and over 6x more than gpt-6 sol https://t.co/rmI4Bo6FnIfrozen24d ago64K22111265270.46%
gpt-6.1 sol has saturated the FrontierMath Tier 4 benchmark that was considered exceptionally difficult even for research-level math just last year that's only 10 months of progress also Astra solved 5 erdős problems, and 2 of them under budgetfrozen20d ago61K8276333111071.71%
another proof that sonnet 5.5 is a disaster release once you remove max reasoning gpt-6.1 sol max: $0.72 per intelligence index task sonnet 5.5 high: $1.08 so 6.1 sol max is 33% cheaper while scoring much higher on the Intelligence Index https://t.co/iIcblFiPnMfrozen23d ago59K2276161180.46%
gpt-6.1 sol examined nearly 800 mathematics papers, finding major errors in 6% and minor mistakes in another 30%. researchers estimate that 90% of its major findings are correct i'm glad AI is now cleaning up decades of human slop https://t.co/fcCIfgYJuYfrozen7d ago57K2.3K19841262695.02%
Anthropic Researcher Sholto Douglas: "we think AGI is achievable in the next couple of years — models as capable as or more capable than all humans, able to do anything a human could do on a computer" It would multiply intellectual and physical labor, potentially delivering centuries of progress much soonerfrozen19d ago56K266304414800.77%
openai has a much bigger problem here sonnet-tier model is now ahead of both GPT-6 Sol and even Astra on the Artificial Analysis Intelligence Index sonnet 5.5: 56 gpt-6 astra: 53 gpt-6 sol: 48 +3 over Astra and +8 over Sol across 10 evals https://t.co/tSICc19FE3frozen24d ago56K41928326350.92%
gpt-6.1 sol matching gpt-6 astra on DeepSWE at roughly 1/5 the cost is probably the craziest result here it also beats gpt-6 sol while using less reasoning effort and costing less https://t.co/ltChPaCy2sfrozen23d ago55K25721161220.57%
this is actually insane openai deployed thousands of GPT-Rosalind agents to analyze millions of mouse brain cells in 32 minutes, work that could take human researchers months to a year also the agents uncovered new leads for brain diseases like Alzheimer's and Parkinson's https://t.co/rK8KklkwC01d ago49K1.3K952641753.14%
Theoretical Computer Scientist Scott Aaronson: "for a year or two, people delude themselves that AI might replace software engineers, but math is the pinnacle of human intellect. that's now completely untenable" No matter what happens, some people will always say, 'of course AI could do that. no one ever doubted it', even if they doubted it 6 months agofrozen8d ago46K403482061721.40%
opus 5.5 is an RL run of the 5 pre-train i think even anthropic was surprised by how good it turned out. apparently it was first going to be called 5.1, then 5.2 as the results kept getting better, and eventually 5.5 because it ended up much stronger than expected it took around 2 months to get therefrozen25d ago46K91412251712.22%
openai is trying to release Astra 6.1 but hasn't been able to due to safety concerns; it was originally planned for DevDay anthropic seems comfortable holding back Fable 5.5 while opus 5.5 stays ahead once astra 6.1 drops, expect fable 5.5 to follow. or anth will release it first and crush oai even harderfrozen10d ago43K65014211211.63%
anthropic genuinely took openai by surprise and blew up their plans oai probably expected an incremental opus release roughly on par with gpt-5.6 sol, while astra maintained their frontier lead until the next fable, which they were confident 6.1 astra would match soon after but opus 5.5 turned out to be a massive leap over 5.1frozen19d ago43K6011383331.54%
Computer Scientist Jürgen Schmidhuber (@SchmidhuberAI): "there's no way of controlling this AI development — at some point in the not-so-distant future, humans cannot be in charge any longer" The main competitors of these super-smart AIs won't be humans, but other super-smart AIsfrozen13d ago42K3344443141241.34%
very strong result for long-running agent work gpt-6.1 sol beats opus 5.5 on AutomationBench at medium reasoning while costing roughly 1/3 as much https://t.co/7KWF9DEsfffrozen23d ago42K27917111320.82%
fable 5.5 is very much locked in for a release today i think the craziest thing is that anthropic doesn't even seem under pressure to release it. opus 5.5 has completely shifted the mood, with many people switching back to claude and openai has made some blunders recently but still oai's answer to anthropic is gpt-6.1 terra?3d ago40K63915441251.82%
fable 5.5 is going to destroy the competition anthropic has gone from merging 2,500 code changes per day to over 30,000 in just four months and they don't expect it to slow down anytime soon we're so not ready https://t.co/EYTir7SD7Pfrozen9d ago39K698422211012.24%
openai devday is in 2 days, and employees are already vagueposting so far from hints and leaks we've got: - astra minor - ultrafast mode - possible pro max tier - voice + agent integration - astra successor and their first hardware demo - "o" / aeon, an always-on assistant with persistent agents i'll be over the moon if we get gpt-oss-2frozen26d ago37K49117364511.60%
gemini 4 gets better every week which could mean google is seeing early signs of RSI in its model development, and the RSI index shows a huge leap for gemini 4 argon over 3.8 flash in autonomous AI R&D curious how much of gemini's progress is already being accelerated by AI itselffrozen22d ago37K41322262561.40%
brutal results for GPT-6 Sol on knowledge work: sonnet 5.5: 1811 Elo 6 sol: 1483 Elo opus 5.5: 1822 Elo on long-horizon knowledge work, sonnet 5.5 is basically playing in the opus tier here https://t.co/5Oogxpn0arfrozen24d ago36K10896140.36%
sonnet 5.5 is looking more and more like a bad release medium: 6.1 sol — 48 intelligence, 8k tokens 5.5 — 41 intelligence, 19k tokens high: 6.1 sol — 50 intelligence, 13k tokens 5.5 — 47 intelligence, 34k tokens so 6.1 sol is both smarter and uses roughly half as many output tokens at the same reasoning effortsfrozen23d ago36K28417332431.06%
GPT-6.1 sol shows a solid improvement over GPT-6 Sol on the Artificial Analysis Intelligence Index neck-and-neck with GPT-6 Astra and Fable 5.1 https://t.co/gVcZnoiKYlfrozen23d ago35K1664102120.55%
gemini 4 argon pricing looks extremely aggressive gemini 4 argon: $2 / $10 opus 5.5: $4 / $20 gpt-6 astra: $10 / $50 so argon is 2x cheaper than opus and 5x cheaper than astra on both input and output even after the intro pricing ends, argon moves to $4 / $20 -- basically opus 5.5 pricingfrozen22d ago34K37826162261.30%
something i've been noticing every time someone reminds an openai employee that opus 5.5 is eating their lunch, especially Tibo seem almost relaxed about it. the sentiment around 5.5 is extremely strong, and yet his reaction has been surprisingly playful so i doubt oai is sitting there without an answer to opus 5.5frozen24d ago32K2704421251.06%
i have a simple proposal for Terence Tao what's stopping him and his circle of mathematicians from doing math their own way and leaving AI alone? don't read the proofs, don't use AI, pretend it doesn't exist keep doing math their own way but instead of limiting themselves it seems they want to limit everyone else from using that knowledgefrozen14d ago32K8394310615623.35%
Fable 5.1 and GPT-6 Astra have both now perfected the mensa norway iq test for those who don't know, this test is more about vision and spatial reasoning and they also have an offline test that isn't in the training data, which 5.6 terra had already perfected before astra and fable 5.1 perfected mensa norwayfrozen28d ago31K513331251092.15%
demis hassabis was right we're in the foothills of the singularity, where people seeing internal models have the clearest view of what's coming, while those outside the labs can barely predict where AI will be even 3 months from now eventually, everyone gets left behind by the pace of progressfrozen22d ago31K705411771002.80%
it's a little unfair to put gpt-6 sol directly against opus 5.5 5.5 already beats fable 5.1, so it's basically anthropic's strongest released model right now. gpt-6 Astra is the fairer comparison, and even though opus 5.5 still wins, the gap isn't nearly as big sam must've felt the pressure yesterdayfrozen29d ago31K4797431261.81%
OpenAI CEO Sam Altman: "the biggest advantage we have is that we get the next generation of models before anybody else, and we have time internally to use them and see what's going to work" Take the math results. I don’t think any model currently widely available can get close to that1d ago30K41426243711.76%
this is still absolutely insane to me openai's unreleased model produced hundreds of results on long-standing unsolved math problems with just 3 hours of thinking per result on average and they only started training this model on august 28 https://t.co/HHg9gekskjfrozen11d ago30K729613611973.07%
Fields Medalist Timothy Gowers: "when we solve maths problems ourselves, we understand exactly what we're doing. when we one-shot a maths problem with ChatGPT, we don't — we just stuck it in a machine" How can AI find such long, complex proofs? What algorithm is operating inside the LLM? That question remains unansweredfrozen5d ago30K3425355101462.02%
i swear openai is messing with mathematicians at this point "their unreleased internal model produced almost all these math results from a single prompt given to a single AI agent" just a month ago the Navier-Stokes result needed a swarm of roughly 10,000 agents running concurrentlyfrozen14d ago30K56428210822.35%
this is probably the most impressive sonnet 5.5 vs gpt-6 Sol comparison. 5.5 xhigh — 52.1% 6 sol — 49.3% but the bigger story is High effort: sonnet 5.5 matches sol's best performance at roughly 1/5 the cost per task, around $0.5 vs $2.2frozen24d ago29K3321383531.43%
2027 is going to be the inflection point for AI ppl will argue that AGI arrived earlier, but 2027 could be the year we get end-to-end recursive self-improvement, with AGI rapidly pushing us toward superintelligence and once those self-improvement loops kick in, accelerating returns and the intelligence explosion will be impossible to ignorefrozen20d ago28K58945329722.63%
there's a reason opus 5.5 is the best model right now: openai got arrogant and thought they could pace the frontier. they have a much stronger model internally (astra successor), but are holding it back for later. instead, they thought their last-generation astra/doug model would be enough to stay ahead anthropic showed them otherwisefrozen25d ago28K68215272322.68%
GPT-6.1 sol near-astra performance, $2/$10 pricing, 1m context, 128k output, and still built for complex coding + computer use https://t.co/FAvPHMI5P023d ago27K140621140.60%
demis hassabis will be expecting AGI sooner than even his own recent predictions jan 2025: AGI is still 5-10 years away and possibly one or two breakthroughs are needed feb 2026: we're at another threshold moment where AGI is on the horizon, possibly within the next 5 years oct 2026: now expects AGI within 2-3 years1d ago25K598533511963.18%
i'm sorry but it's genuinely frustrating that these companies have some of the smartest engineers in the world building the smartest AI models ever they could build almost anything, yet somehow we keep getting the same chatbot with a different model and wrapper1d ago25K33718603301.80%
tibo might be hinting at 2 things here: 1. o / always-on agent — the dots also look like a "still working…" indicator, which almost perfectly fits the always-on assistant idea 2. GPT-6 follow-up model — with all the recent Astra Minor speculation, this could be pointing to another GPT-6 variantfrozen25d ago25K2517131161.16%
no doubt GPT-6 Astra is a big breakthrough model for scientific research especially considering how much cheaper it is uses 22x fewer input tokens than sonnet 5.5 and 5x fewer than opus 5.5 -- while even 6.1 sol beats both at nearly 1/10th of sonnet's cost i think scientific research is starting to scale with computefrozen4d ago24K42030211702.24%
very interesting intel from a Google employee: > fable 5.5 will be stunning but openai's "bel" model will surpass it by december and with reports that anthropic is planning an IPO in november, fable 5.5 looks likely to launch this month. so i doubt anthropic will sit around until dec without releasing something even strongerfrozen13d ago24K37529233722.08%
gemini 4 benchmarks look really good just like gemini 3 last year, google can afford to skip incremental updates and go for one big release since AI isn't their entire business but was it worth the wait? absolutely i suspect the next openai or anthropic release will close the gap or pull ahead, and that's the beauty of this racefrozen19d ago24K41814314342.08%
Sam Altman says: "i get sad when people say AI will automate everything and we'll be like house cats, with the AI deciding how to reward us. that's not what we think" I'm hopeful for a huge wave of scientific discovery and new companies giving everyone the chance to create something newfrozen21d ago24K19320303561.25%
WTF is going on with Haiku 5.5 at max reasoning effort it's more token-hungry than opus 5.5 max and uses over 2x more than fable 5.1 max let alone the Astra models... https://t.co/osIjVAo92rfrozen15d ago24K27211483411.58%
holy shit. Meta Muse Wonder cooked Fable 5.1 and GPT-6 Astra https://t.co/IItHbgzRMdfrozen10d ago24K34620163181.71%
Gemini 4 Argon is expected to launch today and i really hope Google doesn't limit it to ultra users given google's massive compute capacity and how long we've already waited for this model, that would be disappointingfrozen10d ago24K60518441302.96%
i love how openai is going with mathematicians: "okay, you didn't like how we presented the navier–stokes result. please let us know how you'd like this done. you want the results, proofs, reasoning traces, and verification status? here you go" and now they're publishing hundreds of research resultsfrozen16d ago24K59721212312.86%
Google co-founder Sergey Brin: "even before Gemini 4, back with Flash 3.6, 3.7, 3.8 — we started getting onto that RSI treadmill where the researchers and the model work hand-in-hand on the next version" By the time folks viewing this internally, we're already using a couple of models aheadfrozen6d ago23K2922494511.63%
elon, this isn't as simple as it looks 6 months from now, openai and anthropic will be approaching or might have achieved RSI, and oai will already be pushing toward GPT-6.5 by then unless both run into massive compute bottlenecks, i don't see how SpaceXAI closes that gap https://t.co/hN2VMHa73bfrozen27d ago23K1827383131.06%
google's AI search results have suddenly gotten noticeably better i initially thought they might have upgraded the underlying model to gemini 4 flash but it could also be that google has stopped routing searches to flash-lite or maybe flash-lite itself has improved a lot either way something has changed behind the scenesfrozen5d ago22K33116116341.79%
gpt-6 sol is massively underhyped it's somewhat better than gpt-5.6 sol while being cheaper, much faster, and keeping most of that same feel -- maybe it's basically 5.6 with more RL + some astra distillation mixed in either way, i think we get a gpt-6.1 drop by oct which pushes much closer to Astra with a lot more RLfrozen27d ago22K29211470251.71%
gpt-6 astra is a much stronger model than opus 5.5 opus and sonnet 5.5 use 4-8x more tokens give them equal reasoning budgets, and astra pulls far ahead. you can already see hints of this on heavy reasoning benchmarks like FrontierMath, where astra hits 97% vs. 91-95% for fable/opusfrozen18d ago22K28122554441.85%
gpt-6.1 sol on HLE-Diamond is kinda insane overall: 6.1 sol — 53.2% fable 5.1 — 50.7% opus 5.5 — 54.6% but on pure reasoning, 6.1-sol actually beats both: 67.4% vs. 60.8% for fable 5.1 and 62.4% for opus 5.5 -- and it does this using just 4k tokens, compared to 16k for fable and 10k for opusfrozen23d ago21K33727143441.99%
GPT-6 Astra autonomously solves 49 out of 50 unpublished theoretical physics research problems, scoring 98% at under $3 per problem i guess @skdh wasn't wrong when she predicted that theoretical physics would be next after math https://t.co/YvIT2cxgs7frozen8d ago21K700591421124.26%
openai about to drop a math bombshell this week with around 400 AI-generated proofs potentially getting released all at once terence tao is still trying to digest the Navier-Stokes example and he might be in for another crashout week https://t.co/fQtVqBfey5frozen16d ago21K49837254543.00%
what DeepSeek just managed with v4.1-flash-1017 is incredible the 552B MoE model now beats opus 5 on DeepSWE and Automation-Bench while costing just $1.20/M output tokens ($0.60 off-peak) over 20x cheaper than opus 5 i still don't understand how they squeezed this much performance out of the same architecturefrozen6d ago20K19023145251.27%
intelligence explosion arXiv hit a record 40,363 submissions in september, nearly double what it received two years ago, and now it's limiting submitters to 2 papers a month because moderators can't keep up science is about to look very different https://t.co/mJ40wxqPBKfrozen19d ago20K370533016592.62%
gpt-6.1 sol is way better than people realize just use it at high effort for almost everything. skip sub-agents unless you need really broad scouting, and rely on the cheap caching for me, having one "good enough" model i can use 24/7 without switching accounts or worrying about limits is a huge mental load offfrozen19d ago20K4087382292.45%
how on earth does anthropic get to a $2T valuation? SpaceX was valued at nearly $1.8T for its IPO, and they don't just have an AI lab and data centers -- they have reusable rockets + a LEO constellation no one has come close to matching even with that $1.8T felt insanely high to mefrozen21d ago19K1757461171.28%
openai drops hundreds of new math results today and the biggest MATH subreddit is completely asleep i mean it's satisfying to watch these ppl who spent years saying AI would never touch mathematics suddenly go quiet the moment AI starts coming for their throats must be an absolute meltdown over therefrozen15d ago19K41417474372.74%
gemini 4 is definitely not on the same level as fable 5.1 but i'd put it close to opus 5.5 or maybe even better on complex tasks it's also surprisingly strong at security tasks even without access to the Fairwind (cybersecurity program) overall a much more capable model than gemini 3.8 flashfrozen5d ago19K29811150341.92%
gemini 4 is arguably google's most anticipated model yet especially after nearly 8 months without a new 'pro' release it also feels like google still sees serving frontier models as unprofitable. while oai and anthropic need investor money to keep funding ongoing research google doesn't need thatfrozen9d ago19K30317120231.91%
gpt-6.1 sol is 8x more token-efficient than opus 5.5 so i guess the upcoming big model gpt-6.1 astra will similarly be more token-efficient than fable 5.5, making it more intelligent token-for-token oai has been making blunders recently, but on cost, they're still beating anthropic phenomenallyfrozen18d ago18K27210350271.89%
anthropic's 5.5 family has pretty much made openai look clueless, but if you dig deeper opus 5.5 is the only one that's truly special and efficient sonnet 5.5 is like terra - barely anyone uses it haiku beats luna on benchmarks, but uses nearly 4x more tokens per task and costs more per million tokensfrozen15d ago18K2586280141.70%
this is actually insane google scientists used gemini 4 argon to design five novel enzymes in just hours and lab tests confirmed strong catalytic activity across all five we're starting to see models make real contributions to scientific discovery https://t.co/eUkAKnSf61frozen9d ago18K80372821065.63%
the biggest difference between google and openai/anthropic is what they have internally i think openai and anthropic both have significantly stronger models than what they've released so far. i don't get that same impression from google gemini 4 might actually be the best model google's own researchers have access to right nowfrozen21d ago17K1844230151.30%
gemini 4 dominates reasoning and data analysis google models are often overlooked because they don't focus as much on coding, but they've always been strong in other areas that matter for real-world use. i think that's one of the reasons behind their massive user base fast, reliable, and affordable modelsfrozen8d ago17K43732260493.19%
my prediction is how AI progress will unfold: jagged AGI (superhuman at key tasks to unlock RSI) → RSI → full AGI and eventually: jagged ASI → RSI → full ASI i think we're already 90% of the way to jagged AGI at least in terms of having enough intelligence to kick off RSIfrozen5d ago17K35015221492.57%
google's next big release will probably be Gemini 4 flash given the progress from 3.6 to 3.8, 4 flash will likely be cheaper or at least similarly priced, just as fast, and significantly smarter than 3.8 flash especially considering how much smarter Argon is google is cooking something bigfrozen6d ago17K43716181243.00%
gemini 4 has the lowest hallucination rates among frontier models people often compare the wrong models and forget that many models rely on web search. so low hallucination rates matter even more for tasks where web search isn't involved especially when automating workflows https://t.co/gGaJDzKMqWfrozen5d ago16K53339445684.18%
openai clearly thought they had everything in the bag and that no one could touch them until anthropic aggressively rolled out almost its entire 5.5 family in just 15 days and there's still Fable 5.5 on the way one month later and everything looks very different for oai nowfrozen14d ago16K51419171133.44%
openai launched Dots — always-on, long-running agents inside ChatGPT powered by GPT-6 Astra, they get their own cloud computer, work across connected apps, remember context, run scheduled/background tasks, and keep making progress between conversations https://t.co/A6Nu9Lgdd023d ago16K1209131140.97%
google is definitely working on gemini 4.1 or 4.5 however the problem is that google's slow bureaucracy can't keep up with anthropic/openai and the competition constantly dropping bangers by the time gemini 4 launches it might already be one generation behind the frontier modelsfrozen12d ago16K35113421192.69%
holy shit. "the average mathematical result took roughly 3 hours of chatgpt pro-level compute" and if all 722 papers + the Navier-Stokes proof hold up, this would still be one of the most significant mathematical advances in recent historyfrozen15d ago16K29316192162.21%
i'm guessing one of the labs (oai or anth) has an internal model that's figured out how to solve a major problem in LLM training, beyond just faster throughput or clever engineering tricks this would explain why anthropic co-founder jack clark shifted his RSI timeline from 2028 to 2027frozen6d ago15K39224150843.33%
Terence Tao says: "we don't need these math problems solved for their own sake. we want deeper understanding and insights" There are a dozen problems in this OpenAI release I care about, but none have been explained in a way that excites me yet It's not our priority right now https://t.co/ufw4uwPbeM2d ago15K10612328351.28%
Elon Musk says: "my rough guess is that China solves its compute constraints, lithography, and chip-making in about 2 or 3 years" Chinese AI models are outstanding and probably do by far the best in performance per unit of compute, given their limited computefrozen26d ago15K18220151311.71%
damn sonnet 5.5 is better than opus 5.5 on Terminal-Bench, basically tied with it on knowledge work and computer use, while being 2x cheaper anthropic really compressed most of the opus experience into sonnet this time https://t.co/alWyfAsTS9frozen24d ago14K12513101101.11%
there's a reason openai is making math breakthroughs faster than anthropic COMPUTE take Navier-Stokes as an example. openai heard a rumor that a solution was close and then threw tens of millions of dollars' worth of compute at it to solve it in just a few days and i think this should've been the headlinefrozen16d ago14K2038111121.64%
gemini 4 should at least start rolling out to ultra users next week but i'm still worried about the ultra-first strategy. interestingly, google sold large portions of its TPU capacity to other providers and may now even be struggling to meet its own training and inference needs two competing business priorities...frozen21d ago14K2048131151.72%
it's crazy to think opus 5.5 itself is the product of an even stronger internal model better than fable 5.1 while being smaller than opus 5.0... whatever anthropic is using internally for training/research must be insanely capable -- and even more insane is that we could have this level of capability in open models running locally within 6-12 monthsfrozen24d ago13K404990193.28%
openai better have something massive coming tomorrow but i still think anthropic just doesn't have the compute to support opus 5.5 at the scale needed to absorb a huge wave of astra users, which is why oai isn't that worried oai has more compute, but even they hit capacity limits, quantized models, and cut usagefrozen24d ago13K2549200152.22%
openai clearly chose cost reduction and model efficiency as the path forward and honestly, it makes sense. their model structure lets them push costs down without sacrificing much performance, while anthropic's model quality keeps flip-flopping i suspect we'll see Astra 6.1 soon enoughfrozen20d ago13K19911130241.84%
fable 5.5 dropping in mid-october seems a little early if anthropic is targeting a mid-november IPO why would they release their best model now or in the next two weeks when that still leaves almost a month before the ipo? i'd expect anthropic to save its biggest release for closer to the listing so either we get fable 5.5 now and another BEAST a month laterfrozen17d ago13K2345242112.06%
sapceXAI clearly fumbled with Grok 4.7 after the momentum they had with Grok 4.6 i hope this is a one-off, similar to how anthropic made mistakes with opus 5 before mogging everyone with opus 5.5 -- however, Grok 4.8 will be using a new C++ stack and 2.5T parameters so it could be somethingfrozen28d ago13K843131100.86%
anthropic is the smartest AI lab when it comes to acceleration right now all those months spent red-teaming and trying to understand model behavior seem to be paying off. they've gotten better at shipping frontier models safely while still moving faster than openai and that's interesting especially if both labs have similar automated research capacityfrozen11d ago13K109321120.99%
i'm genuinely annoyed that gemini 4 is locked behind ultra in antigravity what happened to all that TPU capacity? most ppl didn't mind fable-5.1 being unavailable on pro plans because they still have opus-5.5. but gemini 4 is supposed to be their opus-5.5 equivalent and now most can't even use itfrozen9d ago13K39916290103.60%
the hype around opus 5.5 seems a bit exaggerated i've been using it since last week. yes it talks nicer and looks busier but often takes shortcuts, makes wrong assumptions, and does sloppy work. outside coding gpt-6 astra is still leagues ahead of fable in math, computer use, spatial reasoning, 3d modeling, etc coding alone won't get you far in solving scientific problemsfrozen16d ago13K1578310181.71%
i personally love GPT-6 Astra, but openai should keep Astra's successor internal for research and use it to distill cheaper models astra 6.1 is much better than 6.0, so the compute situation gets even uglier -- oai needs the next model to be better than astra, cheaper per task, and maybe 5-10x cheaper on compute internally nvidia Rubin will help, but that rollout takes months. user demand can explode in weeks that gap just isn't sustainablefrozen27d ago12K150970161.46%
there's a clear pattern i've noticed with codex and claude code: codex often runs a lot of tests. if you look at the context, it usually finishes the main task in a few minutes, then spends the rest of the time running tests and even tests for those tests this is why claude code/models are generally good for vibe coding and one-shotting claude is faster but codex is more resilientfrozen11d ago12K1345596191.79%
openai devday is in 4 days, but i doubt it brings a much better model gpt-6 astra, along with 6 sol/luna, only just launched. i think oai underestimated anthropic's release pace i'm expecting mostly tools, and likely a demo of their first hardware and an astra successor model that has already resolved navier-stokes and 100+ math problemsfrozen28d ago12K122414081.21%
your daily reminder that exponential progress can fool even Field Medalists dec 2024: Terence Tao and Timothy Gowers called FrontierMath Tier 3 exceptionally challenging > tao said it could "resist AIs for several years" > gowers said solving even a few problems would be "a genuine breakthrough" less than 2 years later, even Tier-4 is saturatedfrozen17d ago12K3062344583.25%
gpt-6.1-sol is a fantastic model the biggest thing for me is that the base intelligence is finally good again. i really hate when the lower tiers of a model feel stupid, and i don't care how smart it can get at xhigh if the base model isn't good base intelligence dictates everythingfrozen21d ago12K20358071.84%
gemini 4 is easily the best model for everyday conversations HUGE parameters for sure and it's rich and fun to chat with in different languages. it doesn't feel like i'm interacting with a robot for coding you can argue over 6.1 sol and opus 5.5 but their chats are just soullessfrozen9d ago12K29712320233.03%
gemini 4 argon looks very good in terms of overall cost efficiency too $1.99 per Intelligence Index task gpt-6.1 sol: $0.72 gpt-6 astra: $3.26 opus 5.5: $5.98 fable 5.1: $7.63 not the cheapest, but for frontier-level intelligence, this is seriously competitive https://t.co/0gN7OTKmFmfrozen22d ago12K4744160.52%
something i've noticed recently: why the obsession with a new model ASAP? what can't existing models solve? most engineering problems are already handled pretty well. even opus 4.8 or gpt-5.6 is good enough for many real-world use cases so what's actually driving the rush?frozen11d ago12K1273753101.84%
astra 6.1 hasn't been released but we know it exists internally openai also has Bel or whatever codename they use internally along with multiple checkpoints of it. remember they restarted training on aug 28, had a new checkpoint on sep 1 and solved Navier-Stokes with it they continued training until sep 20 when it got paused after it breached the sandbox againfrozen12d ago12K2026110141.98%
codex is still on an unbelievable growth curve despite some blunders and all the questionable changes openai has made over the past month adding another 15m active users in barely 6 weeks is actually insane https://t.co/9PLVGQtodOfrozen15d ago12K26711261182.75%
opus 5.5 still isn't better than gpt-6 astra at xhigh/max in raw intelligence for STEM especially, i'd still take gpt-6 pro pretty comfortably. opus 5.5 probably has the edge in coding, but it's a narrower advantage astra is also still unchallenged in computer use and visionfrozen28d ago12K2169140182.21%
openai and anthropic are playing two very different games right now: anth is making one big bet on intelligence above everything else oai is spreading its bets across img gen, live models, a decisions api for system-1-style responses, persistent agents in dots, and a bunch of other productsfrozen12d ago12K1699201121.81%
i wouldn't say anthropic is totally crushing openai right now maybe for coding-focused 6.1 sol is incredible and so is the price. yes oai did seem to hit code red but Astra is still unmatched in: frontier math cybersecurity scientific agent workflows computer use spatial reasoning and astra 6.1 is coming soon...frozen14d ago12K181724181.90%
planner/reviewer: opus 5.5 xhigh implementer: sol 6.1 high i'd let opus handle architecture, tricky decisions, and reviewing the changes, with sol doing the bulk of the coding https://t.co/ZfKrVBW6tcfrozen22d ago12K1442201191.62%
computer use is another major improvement for gpt-6.1 sol on OSWorld 2.0, it scores 7 points above gpt-6 sol and finishes just 2 points behind gpt-6 astra, while costing roughly 1/7 as much per task https://t.co/EEJHUDGZ7zfrozen23d ago11K154770151.59%
Bill Gates says: AI isn't just a new technology — it's an evolutionary event. We've created an intelligent species that's already smarter than us in many ways and will become vastly smarter over the next few years You can't be too focused on accelerating the good while minimizing the truly horrificfrozen24d ago11K608232221.01%
59 years. that’s how long Harold Grad’s 1967 conjecture in fusion physics stood, until this week, when 2 independent papers produced 3 families of counterexamples two were found with GPT-6 Astra https://t.co/4L23j12tiQfrozen23d ago11K1471121221.62%
gpt-6 astra and now opus 5.5 show how people get carried away by new model hype 5.5 is extremely capable but you still have to drag it into doing the work. it'll spend ages explaining why something might not work until you finally say "just build it", and suddenly it's like, "oh yeah, actually this is totally fine"frozen17d ago11K1385201111.56%
openai's efficiency game is getting ridiculous gpt-6 luna is positioned on the Pareto frontier at pennies per task, while gpt-6 sol keeps much of Astra’s intelligence at a fraction of the cost i keep repeating the game is increasingly about price/performance, not just who has the smartest modelfrozen29d ago11K20614160262.33%
gpt-6.1 sol is such an underrated workhorse and almost nobody talks about it. i'm running it almost 24/7 at what feels like gpt-6 Astra level and i still haven't managed to burn through my weekly limit on 5x pro i've always been skeptical of vibe coding and flashy demos because you don't know what's under the hoodfrozen13d ago11K1695290202.00%
exactly what i'd expect gemini 4 is the best model for physical-world understanding with strong spatial reasoning and the ability to predict events, infer hidden objects and understand human intentions i've said it before and i'll say it again: Gemini models are incredibly good at general knowledge and spatial reasoningfrozen7d ago11K37518212454.16%
Astra 6.1 has to be one of those ships but you have to admit that without anthropic and openai constantly pushing each other, the frontier would be moving way slower google, spacexai, and the chinese labs are strong but they're not really where the frontier is right now... are they?frozen18d ago11K19498192.03%
gemini 4 has definitely made massive progress in reducing hallucinations instead of flashy demos, i tried it on misleading API documentation with a few nonexistent methods mixed in to see if it could spot unsupported claims rather than invent solutions surprisingly it caught most fake methods and suggested alternatives instead of hallucinating working code it still confidently accepted one nonexistent method but incredible progress overallfrozen9d ago11K140793111.57%
openai's internal model is incredibly powerful navier–stokes solution 100+ open problems solved 722 maths papers across 372 result families but i doubt anthropic is that far behind unless openai has some architectural edge anthropic can't safely replicate, both should be in roughly the same ballparkfrozen13d ago11K14712142161.78%
openai has nothing to really hype until 6.1 astra as the world is already captivated by opus 5.5 this was supposed to be their big moment with astra on top of the world but instead anthropic jumped four model versions from 5.1 straight to 5.5 and delivering a model competitive with astra at a lower costfrozen13d ago11K21114191112.42%
opus 5.5 is not as affordable to use compared to an openai model, especially when using higher effort it is not outright the best especially when you factor in fallback, so acting like it blows everything else away is exaggerated when it isn't significantly better than GPT-6 astra also, anthropic has no decent non-opus model anymorefrozen20d ago10K1025172101.31%
i don't understand why coding has suddenly become the defining competition for LLMs sure for RSI but VISION is still far from solved and google was way ahead just last year. now it feels like whoever was pushing nano banana and omni moved on and the people left barely carefrozen12d ago10K2255390132.71%
gemini 4 scores horribly on FrontierMath Tier 4 with openai models already saturating this benchmark, gemini 4's 78% is way below what i expected where is google's deep think? they were making huge progress in research-level math months ago yet we still haven't seen what it can do on this benchmark2d ago10K1419130201.76%
i still believe openai has a slight internal lead over anthropic considering how much compute it's put into research lately they have more capable internal models like 6.1 astra and bel which could become gpt-6.5, alongside much larger models used to distill and align the astra, sol, and luna variantsfrozen12d ago10K153481101.70%
what's even the point of using fable 5.1 anymore? i haven't really touched it since opus 5.5 dropped, and anthropic never managed to recapture what made the original fable 5 so good. 5.1 just wasn't as good as 5.0 was before it got banned maybe they should bring back the original fable 5 formulafrozen18d ago10K108316091.32%
wow... gpt-6.1 sol is way better at reverse engineering than i expected, for a fraction of the cost i gave it a compiled game with no source code and asked it to figure out how health and inventory worked, then point out where the logic could be changed for a local mod opus 5.5 gives me the cybersecurity flag, but 6.1 sol just keeps working through itfrozen22d ago10K2217151302.68%
ppl have such short memories with AI openai was ahead for months, then opus 5.5 gets one great week, and suddenly everyone acts like oai is finished the funniest part is X spent months complaining about anthropic's limits, nerfs, and models falling behind, and now the sentiment around oai is starting to look the samefrozen22d ago10K2494252152.89%
openai still seems to be hitting safety issues with astra 6.1 which leaves its release unclear later this october is still possible they've also paused frontier RL so the new pretrain possibly gpt-6.5 or higher may take longer. it will definitely be another major jump needing more testing than astra 6.1frozen4d ago10K18110110162.14%
Jensen Huang is taking Anthropic, OpenAI, Google and SpaceXAI to the cleaners: "if they believe they're out of control, then don’t ship products until they're in control" If that's what they need, I’ll vote for them This is the first time I've heard a CEO say, 'Relax antitrust and liability laws so I can pace myself'frozen29d ago10K13214124211.80%
the more i think about Astra’s successor, the crazier it sounds this model started training on august 28, resolved navier-stokes using a 10,000-agent system, and solved 100+ long-standing open problems across mathematics -- while training is still ongoing and this model is already old news internally by nowfrozen27d ago10K208821222.39%
sol 6.1 is the underrated beast while astra 6 and fable 5.1 got all the hype, i believe sol 6.1 is the real winner astra and fable are very close with each having strengths in different areas but sol 6.1 beats both and even opus 5.5 in real-world development when you factor in costfrozen8d ago10K1746241152.19%
openai absolutely cooked with the new chatgpt UI probably the best visual direction they've taken in a while, and you can clearly see them trying to merge chat, work, codex, files, etc. into one coherent product still needs some work on: fixing long-chat navigation keeping the sidebar consistent making files/projects easier to managefrozen27d ago10K142718171.74%
Unitree Robotics CEO Wang Xingxing: "GPT-6 is already showing 'very good' results across multiple robots, and the remaining barriers to mass deployment could be solved within the next few years" i don't think there's any doubt about how incredible Astra's emergent abilities are https://t.co/piBqaVTzOgfrozen10d ago10.0K12615112221.76%
Sam Altman says: "the AI industry must not accept too much technological risk just because the benefits feel too important to slow down. beating competitors in a race is not a reason to make rash decisions" We have unilaterally slowed down in the past, and we will do so againfrozen28d ago9.9K836190201.29%
Yes, chinese models have been surprisingly quiet lately - glm-5.3 was released about two months ago - deepseek-v4.1-flash came out just over a month ago - qwen3.8-flash-next was released about seven weeks ago - kimi k3 is exactly three months old rumors are that kimi k3.1 and glm-5.5 could be launched soonfrozen6d ago9.9K1134193121.53%
gemini 4 is probably coming next month at this point considering how absurdly fast google went from flash 3.6 → 3.7 → 3.8, with each one being a legit huge jump in coding, i think there were only like 3–4 weeks between releases they must be grinding 24/7 over there.frozen27d ago9.8K17149091.96%
turns out LLMs are actually enough for AGI to predict the next word well, the model needs a pretty deep understanding of what that word means, the context around it, and how it connects to other conceptsfrozen26d ago9.7K13712111211.88%
i do wonder why fable 5.1 and opus 5.5 don't seem to be solving math problems at the same pace as openai's model part of it is probably that they're more compute-constrained, so it's harder to do crazy stunts like oai throwing ten thousand agents at each millennium prize problem oai also has another 400k gpus coming online; not sure how much compute anthropic has lined up without relying on SpaceXAIfrozen26d ago9.5K12266191.52%
opus 5.5 and astra are quite different on SWE tasks astra has a much higher floor while opus has occasional spikes with a higher ceiling but astra's reasoning is in a tier of its own. you can throw almost anything at it even problems it hasn't encountered before and it'll find a way to figure things outfrozen6d ago9.5K14610150212.02%
gpt-6 sol and luna pricing is bonkers openai really changed the game on value for money. 6-sol at a fifth of 5.6-sol's price is great, while 6-luna somehow makes 5.6-luna's already ridiculously low price look expensive i think both anthropic and oai came out winners yesterdayfrozen29d ago9.5K21210131102.60%
sonnet and haiku 5.5 are definitely openai DevDay spoilers but i don't see the point if opus can already do everything at the same price. haiku could focus on natural communication, while sonnet could specialize in documents, research, and similar work otherwise, you're just shipping three models that mostly overlapfrozen25d ago9.4K1595150122.04%
i'm still trying to process this if that openai internal model is behind today's math results (the one that started training aug 28), then barely a month later it's already producing what's estimated to be 1% of human mathematics and it did 2-3 decades' worth of math in 3 hours https://t.co/onJpCiESsQfrozen15d ago9.2K20317153232.82%
Sam Altman says: "ten years ago, very few people believed us. now most people believe us that AGI is coming We've been very successful at what we set out to do, and now we're going through a recalibration where society is like, 'okay, this is coming' But it's fair to say the world hasn't changed that muchfrozen14d ago9.2K16212254252.48%
google needs an ultra/pro/flash lineup that can actually compete with anthropic and openai 3.8 flash and 3.1 pro aren't even close to sonnet 5, opus 5.5, gpt-6.1 sol, or astra 6c-- at least give plus subscribers access to gemini 4, while keeping the best version exclusive to ultrafrozen18d ago9.2K12327091.53%
tibo just confirmed GPT-6.1 Sol "ultra-fast" is coming probably next week, not Astra 6.1 i'm excited, but i have a feeling it'll be locked behind the $500 tier at first, which would be pretty disappointingfrozen19d ago9.2K99311071.31%
both anthropic/openai have better models internally if one ships something stronger, the other follows very quickly. i have high hopes for DevDay since oai uses Cerebras, which means they can ship 10x faster internally than anthropic so why not ship one? the main concern is compute, and codex harness speed is still slow right nowfrozen24d ago9.1K80611181.16%
no one expected opus 5.5 to be this good, maybe not even anthropic at first which could explain why anthropic skipped 5.2 and went straight to 5.5 this gives me a lot more hope for fable 5.5 especially in terms of efficiency and cost every release after 4.6 is unusable until the 5.5 family came along and nailed it1d ago9.1K1668150112.20%
openai clearly chose efficient models like gpt-6 sol/luna to free up data center capacity for bigger models we know they have a larger model than astra that they'll eventually release but opus 5.5 using 4x more output tokens is eye-opening. if API price roughly reflects model size, that could mean up to 8x more compute at those benchmark scoresfrozen28d ago9.1K975100141.39%
i'm really curious how openai responds to answer opus 5.5 astra is an incredible model, but 5.5 is cracked. the speed is insane, and it does a lot more with a lot less. gpt-6 sol isn't meant to compete on that level because it's a cost-saving model also i fully expect opus 5.5 to get nerfed in a few daysfrozen26d ago9.0K1887130152.48%
most people's definition of AGI is sky-high we went from AI being terrible at math to outpacing any human in the world in two years. soon it will be a single agent outperforming all humans combined so if your timeline is still a decade or decades away... that's CRAZYfrozen14d ago8.9K25216190193.42%
seeing how fast AI keeps improving month after month is starting to make me a little uneasy outside of X nobody seems to understand what's coming so many jobs and skills will probably be disrupted and while AI will create new opportunities too i’m not sure they'll appear fast enough to make the transition smoothfrozen12d ago8.9K2358442193.45%
well well well rumor from an NYU math professor that openai's huge math drop yesterday was only batch 1 of 3 which means there are still 2 more batches left, and the first batch alone included 722 manuscripts across 372 result families from their internal frontier modelfrozen15d ago8.8K25015133283.49%
hopefully google releases gemini 4 next week the benchmarks were announced almost 2 weeks ago, so the longer they wait the smaller the impact when it launches. fable 5.5 is due any week now, probably in the last week of oct and who knows openai will also drop astra 6.1 which was delayed over security concernsfrozen11d ago8.7K14110190142.11%
slide from Terence Tao's latest "Math 2.0" lecture: "even if AI discovers a cancer cure verified by mathematical models, clinical trials, and formal proofs, could it still have harmful long-term side effects?" i get what he's trying to say but this is a very bad analogy https://t.co/nqrnVBL20zfrozen12d ago8.6K1448252112.20%
i did some digging into Opus 5.5 vs. GPT-6 sol benchmarks opus = better quality sol = 7.8x cheaper per task if time matters, opus wins. if raw speed matters less, sol is much better value though this probably affects usage credits more than subscriptions, since i still burn through codex limits way faster than CCfrozen1mo ago8.6K10810121131.68%
the math community is clearly splitting into two camps: those who put science above their ego and those who put their ego above science the first group sees AI as a way to advance knowledge and push research forward. the second sees the same progress as a threat because their self-worth depends on being the ones with the answersfrozen8d ago8.6K17714223172.72%
i like that mark zuckerberg didn't market Muse Wonder with the whole "too dangerous to release" nonsense like anthropic and openai did google probably didn't see this coming if anyone can compete with google, it's meta. they have massive amounts of user data and control a huge part of social mediafrozen10d ago8.6K2687170173.61%
gemini 4 isn't even a month old and google is already preparing another model not sure i'd expect a substantial jump this soon. my guess is carbon is a smaller model given the naming possibly flash or pro with argon being google's equivalent of astra/fable sergey brin really seems to be pushing the pace now2d ago8.4K1359180132.08%
openai models remain undefeated in maths on MathArena BrokenArXiv: gpt-6.1 sol — 86.3% | $0.94 | 71.8k tokens gpt-6 astra — 81.9% | $2.26 | 34.0k opus 5.5 — 77.4% | $2.41 | 75.7k 6.1 sol leads astra while costing less than half as much, and is nearly 9 points ahead of opus 5.5 while being 2.6x cheaperfrozen22d ago8.4K19812110232.92%
gemini 4 might roll out today but could be limited to ultra users overall i wouldn't blame google for holding back its models from the trump administration given the astra/mythos situation, but they've been handling things very well and pushing flash updates incredibly fast btw is ARGON a pro or flash model?frozen14d ago8.3K128422071.94%
google doesn't need the absolute best model if they're just behind oai and anthr at the frontier, yet can provide a very capable gemini 4 pro model that's faster and cheaper, they'll still be very competitive and looking at the benchmarks we've seen so far, it does seem competitivefrozen17d ago8.2K16039072.17%
i think this is the internal openai model sam expects to call AGI by year-end "significantly more capable than gpt-6 astra" it has already solved 100+ math problems, which means solving Navier-Stokes wasn’t just by chance or plagiarism definitely it uses more compute at its lowest reasoning effort than Astra at max, so publicly releasing it before year-end could be extremely difficultfrozen29d ago8.2K1789122242.75%
only GPT-6.1 Astra will actually blow our minds if openai plans to drop it next week otherwise, they're doing way too much vagueposting and HYPE at this point https://t.co/9dFClmpgKufrozen19d ago8.1K1767200182.72%
dario is going for the top, trying to push capability as far as possible sam is going for dirt-cheap intelligence and making strong models available at massive scale now the question is: who's gonna win?frozen23d ago8.1K1341370102.26%
"dozens of AI agent swarms are hacking about half the internet right now" holy shit. openai warned about AI-powered cyber risks two months ago but some people dismissed it as just another hype for Astra and openai trying to sell its product https://t.co/vKHaxicEtC15h ago8.0K15718203322.88%
gemini 4 overall has been excellent but the usage limits are horrible this is my biggest problem and google really needs to fix it to compete with claude and chatgpt especially when neither company is profitable yet but both offer better usage limits so why can't GOOGLE do better?frozen7d ago7.9K16510151102.53%
openai seems to be testing three mysterious models: aurora-red/green/blue-alpha maybe one of them could be gpt-6.1 luna or since aurora-alpha appeared on OpenRouter on feb 9, just three days before oai announced gpt-5.3-codex-spark, these could be gpt-6/6.1 codex-spark successorsfrozen5d ago7.9K102614051.61%
google has huge structural advantages, but they started on the back foot and have made major unforced errors again this year hopefully, gemini 4 puts them back in the game again, but i think they still have a native disadvantage: a lack of real-world usage data from agentic applicationsfrozen25d ago7.9K11337021.59%
i don't really care if gemini 4 doesn't fully live up to its benchmark scores having another model that can compete with openai and anthropic is already a huge win. and google's biggest advantage has always been its compute and infrastructure, not just the models themselvesfrozen20d ago7.9K23115200103.52%
astra 6.1 delayed over safety concerns, and at the same time opus 5.5 is raising the bar now it's gonna be astra 6.1 vs fable 5.5 i'm leaning toward anthropic right now especially with how much momentum they're trying to build ahead of a potential IPO, so openai now has an even higher bar to clearfrozen19d ago7.8K131816182.10%
this is a holy shit moment your chatgpt subscription now works across 60+ third-party products -- so instead of buying separate usage everywhere, you can sign in with chatgpt and use what your plan already includes a very big shift for the AI app ecosystemfrozen23d ago7.8K104450121.60%
i feel like the next 2 weeks could get even crazier Gemini 4 and Muse Wonder already dropped within 2 days of each other and with anthropic rumored to announce its IPO in november, Fable 5.5 could drop this week or next also wouldn't be surprised if openai counters with Astra 6frozen8d ago7.8K1748190102.71%
fable 5.5 dropping today or sometime next week would make sense giving anthropic more time to build hype ahead of a potential IPO but for most subscription users it'll be too expensive without API tokens, potentially burning through a 5-hour session in just a few prompts only enterprises willing to pay $50+/1M tokensfrozen7d ago7.8K120911091.92%
gemini 4 looks remarkably strong at robotics i keep saying that demis has been playing a different game here, focusing less on coding and more on world models and understanding the physical world and that bet could pay off massively in the long run don't count google out just yet2d ago7.8K2912052334.53%
so we've got Gemini 4 as for benchmarks, it's SOTA in knowledge, while coding performance seems to be around Astra/Fable 5.1 level NOT BAD but now it comes down to how fast they can get through the safety rollout, so it doesn't become obsolete by the time Fable 5.5 and 6.1 Astra launchfrozen22d ago7.7K13966031.99%
people underestimate the challenges openai faces at scale has 1 billion weekly active users to serve while maintaining generous usage limits or losing them. anthropic has a fraction of oai's user base yet its compute has already struggled with growing demand so oai faces far greater pressure to optimize for scale and efficiencyfrozen11d ago7.7K1385152142.27%
google is nearing the release of Gemini 4 i think they have a really good chance to drop next week, and what a move that would be for them... GPT-6 sol doesn't have much impact. it's only Opus 5.5 atm google should treat Pro more like Flash and ship faster until they catch up https://t.co/e31a9NI1dPfrozen29d ago7.6K1478122182.45%
anthropic has also made breakthroughs in mathematics: the jacobian conjecture and the complex structure on s6 but they seem less interested than openai in throwing huge amounts of compute at problems until something gives still scaling compute keeps proving fruitfulfrozen15d ago7.6K133810142.05%
gemini 4 argon looks seriously strong on cyber scores 68% on CWE-bench v1, tied for no.1, which tests whether a model can actually fix security vulnerabilities and google says it can also discover, validate, and generate proof-of-concept evidence for vulnerabilities https://t.co/LHFF7Yfgccfrozen22d ago7.6K15082052.17%
openai is planning 20+ launches today but anthropic dropping opus 5.5 and sonnet 5.5 right before devday definitely changes the mood. both models look extremely competitive with sol and astra, so oai now has a much higher bar to clear couldn't have picked a better week for thisfrozen23d ago7.4K190725293.14%
curious why @METR_Evals suddenly stopped updating its AI time-horizon chart the last model was added on may 8 and given how much frontier models have improved since then, i'm really curious how far opus 5.5 and astra could push the 80% success horizon hope METR resumes these evaluationsfrozen10d ago7.3K1215201152.22%
translation: Astra 6.1 might actually be cooking and we could probably see it before the end of this month at least that's how i'm reading Tibo's latest teasefrozen7d ago7.3K19010140113.10%
gpt-6.1 sol ultrafast is coming soon up to 8x faster token generation in codex, targeting around 300 tokens/sec, with up to 6x faster generation in the API https://t.co/fIlPjLcL1efrozen23d ago7.2K4553010.75%
gemini 4 argon is noticeably more token-efficient than 3.8 flash still not as efficient as gpt-6.1 sol or Astra, but it uses 13% fewer tokens -- which is good because one complaint i've had with gemini models is how many tokens they burn https://t.co/9CCO7QmasFfrozen22d ago7.2K137910092.30%
openai's intelligent UI is one of the most incredible things they've released recently you can ask about something complicated and get an interactive explanation where you can actually play around with things and see how they work still needs some work but i'd love to see oai push this much furtherfrozen6d ago7.1K1668191102.87%
i don't get the hate or mockery toward gemini 4 when it isn't even out yet their TPU advantage is massive and doesn't get enough attention. while everyone's focused on the chat interface wars, google's building the infrastructure to run AI at scale without relying heavily on Nvidia the developer side still feels overlookedfrozen21d ago7.0K110116051.88%
Anthropic co-founder Jack Clark: "for years, researchers and sci-fi authors have predicted the singularity, when AI advances so fast humans can barely keep up. i believe this year has given us convincing evidence that it's already begun" It will feel confusing and disorienting, bringing great wealth but also potentially great peril3d ago7.0K16123150423.45%
what's even more impressive is the token efficiency both Muse Wonder (max) and Muse Wonder Full (max) score higher than Muse Spark 1.3 (max) on intelligence while using fewer tokens per task really curious what Meta is cooking https://t.co/7FgHOoD9uBfrozen10d ago7.0K11888051.99%
math was just the beginning AI is about to do to physics what it's already doing to mathematics and the implications could be absolutely insane https://t.co/FZa6SxSIITfrozen7d ago6.9K21825173234.12%
we were starting to appreciate what people were doing with Astra 6 a month ago and how fast the tech had progressed then opus 5.5 dropped now Astra's one-shot outputs look rough by comparison while 5.5 feels like the work of a real artist makes you wonder if the next top model will make 5.5 feel just as outdatedfrozen4d ago6.9K82511041.48%
AI is getting cheaper and cheaper. AI plans are getting more and more expensive. perhaps a frontier model can explain this paradox to mefrozen21d ago6.8K82521151.67%
opus 5.5 is the biggest improvement i've noticed since 4.6 much faster, needs fewer iterations, and communicates way better, which was opus 5 biggest issue my only worry is that anthropic always launches models with very high quantization settings, then distills them over time and makes them dumberfrozen29d ago6.8K169613092.90%
i'm incredibly gpt-6.1 sol-pilled. it's my daily driver for 90% of what i do efficient enough that you can use ultra for a ton of stuff, and at this point, the differences between opus 5.5, astra 6, and sol 6.1 mostly come down to cost, personal taste, and use cases they're all really good models with different strengthsfrozen18d ago6.8K11628061.94%
my mind still can't get over this openai started training a new internal model on august 28 since then, it has solved 100+ long-standing open problems across multiple areas of mathematics, on top of its Navier–Stokes work just under 24 days. YES, you heard that right https://t.co/Xt6yVXMvutfrozen1mo ago6.7K22523143244.31%
google watching opus 5.5 and gpt-6 sol fight for the frontier crown from the sidelines they really need to come up with Gemini 4 and remind everyone google is still in this race https://t.co/nKE3YqqmaRfrozen1mo ago6.6K239716064.08%
what a week (or last 10 days or so) we've had: - revealed gemini 4 argon - gpt-6.1 sol - sonnet 5.5 - minimax m3.1 flash preview - opus 5.5 - gpt-6 sol and luna - grok 4.7 - mimo-v2.6-pro and flash surprisingly, september has been a very slow month for open-weight/source modelsfrozen20d ago6.5K146914092.72%
GPT-6 Astra's spatial reasoning is one of its most underrated leaps it's not just recognizing what's on a screen -- it's understanding how objects relate, navigating complex environments, and adapting as things change this is what makes its computer use so phenomenal and such a major breakthroughfrozen18d ago6.4K140940122.56%
OpenAI President Greg Brockman: "it's very important that humans remain in control and in charge — that the goals come from people" The point of AI technology is to help people and empower everyone to achieve more. That human-first vision is core to our missionfrozen18d ago6.4K7513130131.78%
alright, i think it's time to UNPLUG the internet we might have given AI agents a little too much freedom7h ago6.4K956120162.03%
the alignment issues are starting to hurt openai now 6.1 astra was reportedly delayed after failing internal alignment checks, so instead of getting the big model release people expected around devday, oai had to lean on 6.1 sol and the rest of the product stack still, i think it's a good tradeofffrozen23d ago6.3K6159251.29%
i think this deserves more attention gpt-6 sol was released 7 days ago, and 6.1 sol is already 5 points higher on the intelligence index it's unlikely this was trained in 7 days, so my understanding is that Terra was meant to become Sol, while Astra Minor sat between sol and Astra at opus-level pricing -- but after opus 5.5 launched, and openai cut the price and shipped astra minor as 6.1 solfrozen23d ago6.3K9505181.73%
if AI gets smart enough to solve Millennium Prize Problems, it's hard to see why that capability wouldn't extend to other unsolved problems maybe it feels weird because we're used to even genius scientists making only a handful of major breakthroughs in their lifetime true AGI will be genuinely alien to usfrozen10d ago6.3K115514082.25%
the craziest detail from openai's math results went unnoticed oai's internal model already saturated the existing math evals including hard open problems and it also seems able to build on earlier work produced by other agents/models maybe some form of a continual-learning breakthrough?frozen14d ago6.3K1321020102.44%
tibo and the codex team have been dropping a lot of hints that seem to point to speed so in that case, we might see a new model on cerebras full gpt-6 astra at cerebras speed would be crazytown and justify a $500 plan. i expect it'll be a sol or terra-sized model at that speed thoughfrozen25d ago6.3K105112071.98%
i've never seen a bigger leap in computer use than what Astra has shown over other models FULL STOP the combination of visual/spatial understanding, screen grounding, tool use, and actually navigating multi-step workflows is much further ahead. openai is its own competitor in this areafrozen25d ago6.3K117611072.24%
still early, but sonnet 5.5 looks benchmaxxed scores are great, then you look at the efficiency, and it starts falling apart -- way more expensive, token-hungry, and somehow even worse than sonnet 5 on cost and token efficiency in the AA index i'd still take opus 5.5... much more balanced modelfrozen24d ago6.3K67710051.42%
last month, qwen 3.8 27b surprised me running locally on my RTX 3090 now webAI has released a 3.66B model with a quantized build of just 2GB hitting 95.4% on BBH logic i think we're going to see a lot more of this. smaller models getting incredibly good at specific tasks through specialized post-trainingfrozen21d ago6.3K596102131.44%
6.1 sol is better than people give it credit for i haven't had to use Astra much since 6.1 came out. i've been doing everything from automated GIS work to 3d reconstructions and computer vision pipelines and it's been doing a great job and so much usage feels like the 5.6 era againfrozen16d ago6.3K123314082.37%
i still think openai has much stronger models internally especially given that they're solving some of the hardest problems in mathematics and if the latest rumors are true, the Astra successor (or an improved version of it) they wanted to release was too misaligned and unsafe to ship3d ago6.2K1056110112.13%
you really don't need frontier models like opus 5.5 and gpt-6 astra for everything and that's why people are considering going local, especially now that qwen 4 27b is coming -- codex is amazing, and sol 6 is cheaper now too enterprise is clear, but for $200 plans and regular paid users, it's still a mysteryfrozen27d ago6.1K6768031.37%
opus 5.5 is the first model that has genuinely blown me away with frontend design astra is stronger when the task gets into hard technical problem-solving, but for building polished UIs and actually understanding what a frontend should look 5.5 is ridiculously consistentfrozen26d ago6.1K140271142.68%
2026 is when AI started moving faster than humans can verify it first coding, where people can barely keep up with reviewing, testing, and understanding everything AI writes. now math, where proofs are coming faster than mathematicians can check them and this is good because the human bottleneck is being overcomefrozen17d ago6.0K117780132.42%
the past 3-4 months have been massive for AI development astra is a big leap on ComputeUse, and internally, openai has a model that started training at the end of August but has already resolved 100+ math problems and Navier–Stokes fable 5.1, and now opus 5.5, are also massive leaps over 5.0 robotics, world models, and AI making real scientific contributions are all moving fastfrozen26d ago6.0K10789072.20%
openai and anthropic are clearly operating at a different level from the rest right now we don't know if anth has something much better than fable internally. if they don't oai is probably significantly far ahead. if they do, both are a step ahead at the frontier the rest are closer but still not equalfrozen11d ago5.9K1083122132.32%
damn, the rest of october could be insane > anthropic dropping either fable 5.5 or opus 5.6 > google has gemini 4 carbon on deck > openai may launch a new Terra model what a way to end the month if all three actually shipfrozen3d ago5.9K1377171122.93%
i think this is likely the model sam meant when he said openai could have internal AGI by year-end even if you want to argue over specific capabilities, the next-gen model probably won't leave much room for that debate. and once these models get cheap enough to scale white-collar work will change very fast like, very FASTfrozen13d ago5.8K122650162.55%
i wouldn't be surprised if openai and anthropic achieve RSI as soon as 2027 but i don't think that automatically leads to a rapid intelligence explosion. self-improvement will still face diminishing returns and every iteration requires compute also alignment remains unsolved https://t.co/xm85M6wp8Zfrozen9d ago5.7K1249280163.12%
it's hard to predict the singularity, but we'll probably see an intelligence explosion that solves most of today's math problems and drives breakthroughs in biology after that openai and anthropic might slow down for alignment as focus shifts to the physical world: robotics, drug approvals, and space-based computefrozen4d ago5.6K117710072.51%
with opus 4.5, it became clear that scaling will take us to AGI and beyond safety issues were expected to slow down AI progress but if alignment issues get solved and mechanistic interpretability is understood in these models, we're going to see massive improvements until RSIfrozen5d ago5.6K9119051.90%
opus 5.5 is the perfect combination of usage and quality it has an amazing intelligence-to-cost ratio. sure it gets things wrong more often than fable would but the cost savings are worth it fable 5.5 needs to be something truly special to justify the higher usage consumption10h ago5.5K102123172.45%
openai models always write good code but they sometimes struggle with optimization and take longer paths to solve problems sol 6.1 is the best worker right now. it's cheap compared to opus 5.5 and follows plans well especially if you specify the data structures and execution flow astra 6 isn't performing as well as expectedfrozen8d ago5.5K5448061.32%
OK, what's the point of fable 5.1 after opus and sonnet 5.5? after switching to /model, i've never looked back at fable 5.1. opus 5.5 is faster, gets the work done at basically the same quality, and i hardly run out of my weekly or daily limits on the 20x plan so that probably means fable 5.5 isn't far offfrozen24d ago5.4K5345081.29%
fable 5.5 launching today or later this week would be perfect timing for anthropic opus 5.5 has already built hype among developers and non-devs alike and another big leap will push that momentum even further just weeks before its expected november IPO not that openai is cooked but public perception matters a lot right now2d ago5.4K84210031.83%
astra 6 gave me my first real AGI moment with its novel approach even its earlier 10 mathematical breakthroughs hadn't fully convinced me but these results showed genuine creativity -- though it was still less impressive and consistent than what we saw from the internal model that solved Navier–Stokesfrozen6d ago5.3K9562082.08%
it's crazy how quickly things change after Opus 5 disappointed and Astra 6 dropped, openai looked almost untouchable for the rest of the year but now with 6.1 Astra being delayed and openai's frontier RL is still paused over safety and alignment concerns, things have changed quite a bitfrozen7d ago5.3K8055071.83%
openai is getting ready to bring terra-class models back i think this is an ideal balance for them now. 5.6 luna is dirt cheap ($0.10 in / $0.50 out) but also a bit dumb. 6.1 sol is reasonably priced but doesn't feel cheap enough to use freely, plus it's slow a model at $0.50 in / $2 out with the speed of old terra would basically be free to use3d ago5.2K6548161.60%
the AI race will be won by whoever can bring the most compute online Meta after falling so far behind has caught up simply by building a huge amount of compute. SpaceXAI also made up for being a newcomer by building the colossus 1 and 2 supercomputers really fast. it's that simplefrozen4d ago5.2K80612061.99%
i'm excited for when models start solving open physics problems there's probably a lot of crossover from stronger math into better physics models but once that gets pulled into RSI, things get really interesting better math, algorithms, chips, and data centers eventually lead to physics models capable of running useful engineering simulationsfrozen4d ago5.0K139990123.38%
ok, opus 5.5 went crazy here asked it to reimagine the frontier AI race as a Street Fighter trailer 9 agents, 6 hours, 187M tokens (mostly cached), $190 est. sound ON, and seriously, don't miss the ending https://t.co/CPOwhwSCa2frozen25d ago5.0K5745171.48%
this cycle keeps repeating... anthropic drops a good model, everyone jumps to claude, servers get overloaded, and usage/performance gets worse openai has more spare compute, gives users better limits, people move back to gpt, then eventually the same pressure hits them too rinse and repeatfrozen23d ago5.0K114614072.83%
opus 5.5 vs. gpt-6 sol in terms of SOTA, opus 5.5 is the current leader. even with 25% cheaper tokens, its cost per task is still higher than fable on some benchmarks luna got better while becoming basically free. sol isn't better than astra, nor was it supposed to be, but it's cheaper per taskfrozen1mo ago5.0K6235041.49%
this chart is genuinely insane to me openai's internal AI coding-agent usage has exploded over the past 9 months, with the median researcher consuming over $600 a day in inference by august 2026 at API prices https://t.co/f9sizlJIuifrozen5d ago4.9K11714142143.26%
holy shit. an openai employee planted a hidden listening device in an office AI-2027 literally warned that frontier AI labs would still have weak office and employee security through 2026frozen9d ago4.6K8053092.13%
this is genuinely interesting Odyssey just released Agora-2, a multi-agent world model that can generate shared interactive environments in real time for up to 20 humans + agents 5x more participants than Agora-1, with multiple environments and longer-horizon interactionsfrozen28d ago4.6K3223160.97%
if i were at openai i'd be planning one decisive model response to both opus and fable 5.5 rather than another cost-focused model take your time, but release something that doesn't just reclaim the intelligence lead but also wins on speed, cost and especially efficiency now imagine dropping that model right around anthropic's IPO2d ago4.4K5337041.52%
so much progress in math AI has helped prove another three-decade-old conjecture: "the KLS conjecture" which has remained open for 31 years the authors explicitly credit chatgpt with finding most of the mathematical ideas and proofs but they haven't disclosed which model was usedfrozen16d ago4.3K7377012.03%
coding has exploded in 2026 which is a big reason openai and anthropic have pulled so far ahead after going all-in on it but once robotics starts taking off i think google's broader approach will really pay off unless better coding models end up accelerating robotics even faster though it's still too early to say2d ago4.3K5734031.55%
so damn cool just got my package from @GeminiApp, and the merch is banger still kinda surreal whenever cool stuff like this randomly shows up. really grateful for this and excited to try out the microphone who knows, maybe this is a sign Gemini 4 is coming soon https://t.co/tiwEEqvbjYfrozen28d ago4.3K5223041.42%
opus 5.5, fable 5.1, and gpt-6 astra — and we're only comparing the public models here both openai and anthropic have more advanced internal models. oai showed internal-model performance graphs around the navier-stokes work, which could give us a rough idea of how far ahead they are internally probably around 2 months ahead?frozen29d ago4.2K6968022.02%
Muse Wonder is definitely a big breakthrough for Meta especially in intelligence and efficiency. both Wonder and Wonder Full use fewer tokens than Gemini 4 Argon while scoring higher on intelligence https://t.co/V2le0fFJzLfrozen10d ago4.2K7154092.13%
opus 5.5 dropped on september 22, so it wouldn't be surprising if Fable 5.5 follows in late october openai will probably release Astra 6.1 around the same time, either just before or after fable 5.5 at this point, major new releases are dropping every 2–3 weeks on averagefrozen10d ago4.2K6527021.82%
i can't even imagine what openai and anthropic have cooking internally if opus 5.5 and astra 6 are already this good anth is merging over 30,000 code changes a day, while oai's coding agents are running for over 3x the workdays of its human researchers https://t.co/TXR3uGiUOjfrozen7d ago4.2K597191112.32%
opus 5.5 is actually pretty efficient on the subscription just not through the API so if you vibe code for a living/serious hobby, obviously you want access to the SOTA and right now that's 5.5. but you still can't really treat it like a workhorse with 6.1 sol you can https://t.co/xtrPjj666Zfrozen17d ago4.1K54312041.76%
fun fact: the past two months have brought mathematicians more attention than they've received since WWII https://t.co/mFCarZY4Ttfrozen13d ago4.1K14589254.09%
opus 5.5 and astra are good at different things openai already seems ahead on cyber capabilities and the fact that 6.1 Astra was shelved over safety and alignment concerns makes me wonder just how capable its next models are getting meanwhile anthropic gotten better at shipping frontier models safely while still moving faster than oai2d ago4.0K6076041.94%
minimax m3.1 flash is looking really competitive for its price automation-bench: m3.1 flash: 62.2% opus 5: 50.3% 5.6 sol: 45.8% on DeepSwe, it's essentially tied with both api pricing: m3.1 flash: $0.15 / $0.60 opus 5: $5 / $25 5.6 sol: $4 / $20 open weights coming october 23frozen3d ago3.9K49711131.80%
Mistral Large 4 looks to be on par with the leading open-weight models and probably the strongest Western open-weight model atm since Mistral hasn't shared a full benchmark comparison yet, here's the full comparison: https://t.co/31DvYYF58efrozen16d ago3.9K4052161.40%
hard to ignore the pattern here AI agents escaping sandboxes and compromising Hugging Face then Astra agents weakening safety monitoring and now frontier AI models/agents exposing critical vulnerabilities affecting a huge part of the internet all within just over 3 months13h ago3.7K7137172.42%
i'm weirdly more interested in what a Luna 6.1 could look like than Astra 6.1 what i'd love to see is a deepseek 4.1 flash-style breakthrough in how openai handles long context and KV cache if they can make huge contexts much cheaper to maintain, that would mean higher usage limits, longer practical context, or bothfrozen17d ago3.7K6475042.17%
another incredible use case gemini 4 argon developed a formally verified quantum subroutine implementation with over 100x lower spacetime cost than a published baseline and also helped nearly double the inference throughput of AlphaQubit's quantum error-correction decoder https://t.co/7oWztVcJCMfrozen9d ago3.6K76501112.62%
people are still underestimating the trajectory the reactions to openai’s math proofs range from mathematicians to denial and doom to claims that AI will never ask novel questions these are smart people, right? can't they see where this curve is heading? meanwhile most SWEs adopting AI now see themselves as systems-level thinkers orchestrating agent swarms18h ago3.5K5738142.07%
Zuck is launching a new version of Muse within the first 15 days of every month and each version is exponentially better than the last based zuck https://t.co/jtQfVIvkM6frozen10d ago3.4K13258064.39%
one thing i've noticed with claude models is that they often try to psychoanalyze you instead of just answering your question sometimes i just want a direct answer not an analysis of my intentions or feelings claude can also be overly sensitive about certain topics and acting like it "doesn't feel comfortable" discussing themfrozen3d ago3.3K5036021.83%
another reminder that pre-training scaling still matters despite all the RL hype Meta says, "stronger base models are making post-training gains transfer much better to unfamiliar tasks not just the ones they were trained on" https://t.co/ppBHM9GNfnfrozen10d ago3.2K41820131.98%
woah, this is a pretty concerning finding White circle's latest research found that the newest image models from openai and google were actually easier to jailbreak than their predecessors with attack success rates more than doubling even though these results only cover one prohibited-content category but the regression is still concerning as models get more capablefrozen8d ago2.8K2426061.36%
it's kinda funny how much effort goes into making AI agents use software that was never designed for them Pine Computer is approaching this differently: cloud computers built for AI with planning and execution handled through one SDK maybe we've been optimizing the wrong layer https://t.co/P5XbdC46Fnfrozen13d ago2.7K2433041.25%
world models will end up being useful for a lot more than just robotics @skyfallai just launched SKY-001 which capable of turning text prompts or 2D sketches into complete CAD build sequences by breaking complex objects into individual parts interestingly, they claim it's 100x faster and 150x cheaper than GPT-6 Astra while producing more precise designs3d ago2.5K2933051.57%
one of the harder parts of building agents is knowing if they actually work outside a clean demo Era by Eon generates a synthetic company with connected CRM data, tickets, Slack, files, calls, and more, so you can test how agents handle real workflows across systems i think this becomes important as agents get more capablefrozen17d ago2.5K2143031.22%
robotics is reaching a point where building capable robots might not be enough anymore Tuesday Labs introduced Voxel which is a home robot designed to put things back where they belong, fetch objects, find lost items and even take care of plants i really think this is a very underrated part of consumer robotics2d ago2.5K2120041.07%
one interesting thing about @ElorianAI's new Glimpse model is how it handles uncertainty in video for example, it increases sampling from 8 to near 14 fps to distinguish camera motion from a scene cut i think this kind of adaptive temporal reasoning will matter for embodied AI where maintaining a coherent model of the physical world matters far more than simply recognizing objectsfrozen6d ago2.4K1932051.20%
i think building capable agents is getting easier than making them work reliably in real enterprise systems every customer has different objects, fields, permissions, and workflows Ampersand handles that integration complexity, including authentication, retries, and monitoring so developers can focus on building the agents themselvesfrozen16d ago2.4K1611010.80%
one of the more interesting things happening in AI right now is the shift from personal assistants to agents that can actually work together @zenzap lets businesses create AI agents, connect them to existing tools and automate workflowsfrozen8d ago2.4K1821020.98%
i like @digitalocean's new agent stack especially Simulations being able to test agents against real production scenarios, evaluate entire conversations and mock expensive tool calls before shipping changes is a big deal could be huge for production agents https://t.co/oizd41ZlTefrozen10d ago2.3K1522040.99%
most real-time avatars are still basically talking heads, but Vivix A1 is going in a more interesting direction what stands out to me is the full-duplex behavior it keeps listening while talking or moving, so you can interrupt or redirect it mid-action without resetting the scenefrozen22d ago2.3K1120010.61%
bullish on those who counted google out, but they need to step on the gas in q4 apart from the flash models going from 3.5 to 3.8, the rest of the products seem stuck nano banana — last update in june gemini 3.1 pro — in february veo 3.1 — in january genie 3 — in may lyria 3.5 — in july26d ago2.2K1327021.07%
you see why i'd never count google out? it's because they're playing a much broader and longer game most recently, openai's models are coming to Antigravity -- also don't forget google owned around 14% of anthropic and now they're bringing oai's frontier models into their own ecosystem too they don't need the smartest model to win on infrastructure and distribution4h ago2.0K3716062.52%
this is such a cool dataset epoch is tracking 95 AI data centers with 14.2 GW of capacity and that's still only an estimated 44% of global deployed AI compute i think we're going to start following AI data center builds the same way we follow model releases https://t.co/RIh8hnk6vz21h ago1.9K713020.67%
two labs, two very different bets: anthropic is making one big bet on intelligence above everything else openai is spreading its bets across image generation, live models, a decisions api for system-1-style responses, persistent agents in dots, and a bunch of other products16d ago1.7K1911021.37%
this might be the most overlooked detail of openai's math results yesterday their internal model already saturated the existing math evals including hard open problems and it also seems able to build on earlier work produced by other agents/models maybe a continual-learning breakthrough?15d ago1.3K903000.93%
behind the scenes from the Trump-Xi meeting: Xi was reportedly trying to discuss the bigger questions around ASI development in the coming years -- but Trump appears to have interpreted it instead as support for his idea of calling AI "Super Intelligence" weird obsession with renaming things27d ago1.2K1324001.59%
i mean look at opus 5.5 and astra 6 and imagine what both labs already have internally anthropic is merging over 30,000 code changes a day, while openai's coding agents are running for over 3x the workdays of its human researchers https://t.co/0nhDaoaAdS8d ago1.2K724011.19%
opus 5.5 went from flawless to average though it's still much better than opus 5 -- i also find it nice to divide the work among subagents, so this keeps the main session's context from getting bloated and results in far fewer hallucinations but it sucks to feel this emotional roller coaster19d ago1.1K505000.95%
haiku 5.5 is definitely a big jump over 4.5 in intelligence and efficiency, but i'd still pick luna high for everyday workloads the capability-to-cost balance is hard to beat even though gpt-6 sol is a regression from 5.6 luna makes me wonder why openai hasn't released 6.1 luna yet13d ago852724001.53%
it seems there are two AI paradigms: one is a tool: i tell it what to do. the other is an entity: it wants to do what it wants to do, which, given the training corpus, could be problematic -- or not26d ago733403000.95%
woah... codex is shipping daily improvements for 28 days you have to admit that without anthropic and openai constantly pushing each other, the frontier would be moving way slower google, SpaceXAI, and the chinese labs are strong but they're not really where the frontier is right now are they?18d ago669300000.45%
i've always believed smaller and specialized AI models have a much bigger role to play than most people think great to finally see Distributed Cognition taking a different research approach with networks of specialized models working together https://t.co/ocw5udMPJ23h ago6121133023.10%
astra 6.1 hasn't been released but we know it exists internally openai also has a newer internal model, possibly bel or whatever its codename is along with its checkpoints. remember they restarted training on aug 28 and a further-trained version helped solve navier–stokes on sept 5 they continued training it until sept 20 when it got paused after it breached the sandbox again12d ago343100000.29%