← Back to leaderboard
hqmank

Kai

Verifiedactive
@hqmank

Kai Wang (AI Kai) AI Builder 🛠️ Practical AI Workflows & Coding Tips High-Value Promo Codes Updated Regularly DM for collab 📩

14,396 followers · 257 following · Last polled 4h ago
Perf. Score
43.8
Rank #277
Value Score
—
set rates to compute
Median views
2.8K
p25 2.2K · trailing 7d
Followers
14K
+0.6% 7d
ER (impr.)
1.37%
eng ÷ impressions
ER (foll.)
0.78%
eng ÷ followers
Est. CPM
—
needs a rate
WoW views
-10.8%
▼ falling

Rates & value · est. from 2.8K median organic views

Quote tweet
—
no CPM
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
rate 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 (2.8K), not the mean (8.2K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (2.2K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
Claude is giving away free plushies. Go to https://t.co/GX84b3uQEA, click terminal, then type: /plushies for a random plushie /stickers for a sticker pack Worth a try if you're into it. https://t.co/H86fB46AMwfrozen22d ago2.2M8.7K3943993077.8K0.79%
Andrej Karpathy makes a good point: as AI does more of the work, more of your job becomes reading what it produces. His first tip caught my eye. Ask the model to explain things in ASD-STE100. ASD-STE100 is a controlled English standard built for aircraft maintenance manuals. A mechanic anywhere in the world should read a step once and get it right. The core rules: 1. Keep sentences short 2. One action per step 3. Say who does what 4. Give each thing one name and stick to it AI output breaks that last rule all the time. It says "worker," then "agent," then "executor," and you can't tell if it means one component or three. The full spec is strict. Karpathy asks for "80% of the way," and that works for daily use. Someone turned it into an open-source skill: https://t.co/FFqEehqVl2frozen
21d ago
932K
6.5K
451
70
19
13K
2.11%
🚨Anthropic removes Cowork's local option for Pro/Max tomorrow. New tasks run in the cloud. Existing local tasks stay on your computer. Anthropic says your sessions and files "live with your Claude account and go where you go." Where is my data, and who can access it? https://t.co/bsUSxGdEZKfrozen17d ago470K1.8K99147876370.60%
Tibo's reset button is real. It's not a render. It's an actual box with a silver flip cover and one button on top. Every time Codex limits got reset, people joked there was a button somewhere. Turns out there is one. No approval chain. Flip the cover, press it, done. DevDay is tomorrow. Tibo, feel free to press it a few more times this week! https://t.co/qzGqn4fH3Efrozen24d ago191K6983332131530.49%
Another way to sign up for Muse outside the US: Sign up at https://t.co/E48o5nUo1G with GitHub to get free credits. Ask its cloud browser to open https://t.co/0er64DjLdC, then complete the signup. You can link your Facebook account to skip the card check. https://t.co/bMh1nM7JI1frozen28d ago189K4212910977660.71%
Not in the US? Here's how to sign up for Muse. You'll need a paid Google Gemini subscription with access to Spark. 1、Go to https://t.co/sIEVrHSojp and enter: "Open https://t.co/L1LXDH5xLY in your remote browser." 2、Allow Spark to interact with the site, then take over the browser and complete the signup. 3、If you're asked to verify your age, follow the on-screen instructions. You may need a supported card in your own name. This method may not work in every region.frozen28d ago158K214135863640.41%
Google AI Pro and Ultra users: don't forget to redeem your monthly Google Cloud credits. Pro: $10/month Ultra: $100/month Google just posted an official step-by-step guide. Learn more about how to redeem them below: https://t.co/zw5oMlvaVlfrozen14d ago97K1.1K653031.3K2.57%
Google's student offer just got more interesting. Some users report switching from the free AI Pro student plan to the $8.99/month bundle with YouTube Premium, and seeing their next payment pushed back by over 2 years. YouTube Premium alone makes this pretty tempting. If you already have a verified student account, check the offer at https://t.co/aWSF55oWxm.frozen4d ago97K409132215331.01%
OpenAI changed its ChatGPT Pro pricing. The two tiers are now "Standard" and "More". Standard still costs $100/month. The old "5x Plus usage" line is gone. The card now says "More usage than Plus" and nothing else. Same price, so how much usage does Pro get now? Still 5x? Or 3x?frozen24d ago86K25314413460.42%
The Codex reset button everyone keeps begging Tibo to press. Turns out it's a real physical device. Roughly looks like this. @thsottiaux https://t.co/YzM2nwPUqJfrozen23d ago68K198632270.35%
Day 10 of Gemini 4 Argon. Still can't use it. What's going on, Google? https://t.co/xDT2qAsO5kfrozen14d ago64K1.7K39654332.89%
Tibo says GPT-6 Astra and GPT-6.1 Sol got ~50% faster. I tested it: 20 tok/s → 30 tok/s. The +50% is real. But Claude Opus runs around 80 tok/s and Sonnet around 100. Faster than last week, but still pretty slow. https://t.co/rJqOV2yOWBfrozen17d ago59K25013302190.53%
Tibo has a real physical Codex reset button. I rebuilt it in 3D from the original video, so now anyone can try it in the browser. Pull the handle down and a 10-second countdown starts. Cancel before it hits zero, or let it run and watch the reset go through. https://t.co/U1W7bhC5UNfrozen23d ago53K910122390.27%
Didn't expect Gemini 4 Argon to beat Astra and Opus 5.5 on half the coding benchmarks here. And it leads every knowledge work benchmark in the table. Looks promising, but I trust developers who've used it on real projects more than a benchmark chart. Going to try it myself today.frozen22d ago53K1904220.05%
Every tip in @karpathy's post is gold. I use the HTML one a lot. When I'm learning something new or trying to understand complex code, I ask the agent to build me a visual HTML page. Here I was learning SEO and GEO. The agent turned the key ideas into an interactive page I could click through. I picked it up faster than I would have from reading docs. https://t.co/C4PkQd3kEXfrozen18d ago51K605492109253.12%
Wanted to try Gemini 4 Argon, but can't find it in Antigravity. Anyone got access yet? https://t.co/5IC6MyVRMzfrozen22d ago47K1850260180.48%
Grok, Cursor and X used to be three separate subscriptions. Now it's one. The $30 plan gets you what SuperGrok, Cursor Pro and X Premium+ used to cost together: $90 a month. Mostly coding? Spend your whole allowance in Cursor. Mostly chatting or making videos? Spend it in Grok. It's one pool. Already paying for any of them? Nothing changes for now. Keep your plan, or switch and get the unused time refunded.frozen10d ago46K269102821130.91%
Claim 1 billion free Muse tokens within 48 hours of signing up. Go to Settings → Redeem invite code. You and the person who shared the code each get 1 billion tokens. Find a code by searching X for "Muse invite code". https://t.co/Py8U1A0cdWfrozen28d ago46K2651213433221.59%
📈 Your Google AI subscription got more valuable. Antigravity added Claude Opus 5.5 and Sonnet 5.5. You need paid Pro or Ultra to use them. The free student Pro plan doesn't include them. GPT-OSS 120B leaves Antigravity on November 2. After that, student accounts only have Gemini models. Add Gemini 4 Argon to the paid plans, and they get hard to beat.frozen20d ago45K2657327920.89%
I just updated ChatGPT to the latest version, but I still don't see Auto-review. https://t.co/zbjP6Xl3aRfrozen16d ago42K2063171350.62%
Quick heads-up: Not every Google AI Pro subscriber gets access to Gemini 4 Argon in Antigravity. These users are currently excluded: - Students with free AI Pro plans - Users who got AI Pro through a Pixel purchase - Jio-bundled AI Pro subscribers - Users on the older Antigravity IDE - Ultra for Business users If you have Pro but can't access Argon, check your subscription type and Antigravity version first.frozen5d ago39K18012294510.71%
Tibo says they've optimized it to 50 tok/s. I'm getting 26 tok/s with GPT-6.1 Sol in Pi, high reasoning effort, no Fast Mode. Am I missing a setting, or has the update not reached me yet? https://t.co/yyIMjQNPw8frozen17d ago37K3545493181.16%
DeepSeek finally has its own desktop app. It's in Preview for now. Want to try it? Links below 👇frozen27d ago36K1675123580.67%
DeepSeek released its official Harness desktop app for macOS and Windows. No Linux version. You can sign in with your DeepSeek account or use an API key. I've been using it for a while and like the clean UI. Their API pricing doesn't feel competitive to me anymore. Hope they launch a Coding Plan soon.frozen23d ago35K488282252072.12%
Tibo handed us a banked reset. Claude matched it the same day. Scoreboard: 1:1. Round 2: first one to kill the 5-hour limit gets my whole weekend.frozen1mo ago35K5734121151.71%
Kimi hired a 16-year-old to work on frontier AI models. Less than a year later, he's leaving with his name on one of their core architecture papers. He's 17. No PhD. At many labs, his resume might never have made it past the first screening. For years, PhDs and established labs have dominated frontier AI research. Without the credentials, you rarely got a shot. That's getting harder to justify. Papers, code, and open weights are more accessible than ever, and AI makes it easier to work through research you might otherwise struggle to understand. The work still has to hold up. Kimi gave him the chance to prove he could do it. I hope more labs do the same.frozen7d ago34K47715311391.88%
Just realized Google AI Pro's monthly Cloud credits are valid for a full year, not just one month. I forgot to claim mine for several months, and I can still redeem them. Worth checking if you've missed any.frozen6d ago33K35016723282.12%
Gemini 4 Argon is now rolling out in Antigravity for paid Google AI Pro subscribers. Pro trials, Free, and Google AI Plus are excluded. Here's the catch: Argon gets its own weekly-only quota. No 5-hour refresh. The good news? It's separate from the Gemini 3.8 Flash bucket, so using Argon won't eat into your Flash quota.frozen7d ago33K1584101190.59%
If you pay for Google AI Pro or Ultra, don't miss the Colab benefits included in your subscription. Pro gives you access to A100 80GB GPUs and 200 compute units per month. Ultra adds H100 access. GPU availability varies. Free trials, Jio offers, and student offers don't qualify. Link in the comments 👇frozen12d ago32K308122022741.93%
Was just running Gemini 3.8 Flash in pi agent. Looked down and it said 1440 tps. Is Ultrafast just on by default now? https://t.co/DQirdbTkrCfrozen22d ago31K2542192451.03%
5 ways to try signing up for Muse outside the US: Open https://t.co/L1LXDH65Bw using: 1. Hyperbrowser (https://t.co/ZWuuuNqMbB): start a cloud Chrome browser. 2. Lexmount (https://t.co/p7HFhEjkAg): start a cloud browser. 3. Gemini Spark (https://t.co/sIEVrHSW8X): ask it to use its remote browser. Requires Spark access. 4. Grok Bot (https://t.co/XvMlMILplG): ask it to open the page on its cloud computer. 5. A US residential IP: use your own browser. Finish the signup and any verification yourself.frozen25d ago30K15493752421.49%
Considering ChatGPT Business for two people? Right now it's buy one Standard seat, get the second free for 48 months. Fill out the short form, and if your application is approved, you'll receive a link to redeem the offer by email. I went through the process myself. The redemption page showed 1 free Standard seat for 48 months. Form link if it's useful: https://t.co/qFfmx0q7fbfrozen24d ago26K12281001591.14%
Antigravity models inside Pi, without the reverse-engineered OAuth hacks. Most ways of plugging Antigravity into other clients go through the private Cloud Code Assist login. That's how people get their main Google account flagged. pi-antigravity-acp-provider takes the official route. It talks to Google's own Antigravity ACP server. You keep the full Pi workflow, and Antigravity runs the model underneath. Setup: pi install npm:pi-antigravity-acp-provider Then in Pi: /login → Google Antigravity (ACP) → sign in through the browser /model → pick any antigravity-acp/* modelfrozen18d ago26K323261903292.67%
Gemini 4 Argon launched again. Google AI Pro and Ultra subscribers still can't use it. Google is only giving access to cyber defenders and trusted testers. Is it Mythos 5.1-level powerful, or what's keeping it from paying users?frozen16d ago25K2386201171.14%
Codex limits reset today. Can't get excited about it at all... Anyone else feel the same?frozen9d ago24K338480231.74%
If you're paying for Google AI Pro or Ultra, check Colab. Your plan now includes monthly compute units. Worth trying for a quick eval or a small model experiment. You can run jobs from your terminal with the Colab CLI. Stop the runtime when you're done. Idle GPU time eats into your credits too.frozen29d ago24K222450931.34%
GPT-6.1 Sol has felt slow to me. Tyler checked his logs to see how slow. He had an agent analyze about 65,000 calls from his coding tools, comparing answers of similar length and measuring the wait from sending a request to getting the full response. In his data, GPT-6.1 Sol was the slowest of the 15 models he used. Even Fast mode only matched regular GPT-6 Sol.frozen18d ago24K21216146301.17%
You can use DeepSeek-V4.1-Flash free in WorkBuddy until Oct 31. https://t.co/co8kRICBOzfrozen12d ago23K358182532022.64%
Codex users are leaving because of the model. Optimizing product details won't change that. Ship a model that can go head to head with Opus 5.5, and ship it soon. Speed is the other gap. In my tests, GPT outputs around 20 tokens/sec, while Claude does 60-100. Codex needs to at least catch up. Unless OpenAI has decided to stop competing on models.frozen18d ago22K4256512162.25%
Muse gives every account its own Linux VM. I ran a few commands inside mine: 2 cores, 8GB RAM, 100GB disk, and I'm logged in as root. So this isn't a chat box pretending to run code. The agent gets a real machine where it can install packages and run scripts. https://t.co/1BbOPeV7oxfrozen26d ago21K220194141281.93%
Tibo might be the most active OpenAI person on X. He's always replying to bug reports and resetting Codex limits. You'd think he worked in DevRel. But according to WIRED, he leads OpenAI's core products, including ChatGPT and Codex, and reports directly to Greg Brockman. https://t.co/6MwCywgH3Lfrozen5d ago21K1643212110.94%
Google AI Pro and Ultra now come with Colab perks. Pro gets about 200 compute units a month. Ultra adds background execution and Premium GPU access. Some people told me they haven't gotten it yet. That's because Google is rolling it out in phases across supported countries, not to everyone at once. Google One trials don't qualify. On family plans, only the plan manager gets it. You also need to be 18+. To check, sign in with your subscribed account and open Colab's resource panel: https://t.co/fH9Kcpm6sqfrozen22d ago20K123751691.02%
A Google product lead says Gemini 4 Argon is coming to Antigravity soon. Google says paid API customers and Ultra subscribers get it first. Can it beat Opus 5.5 in Antigravity? 👀 https://t.co/vXjGLzA2NKfrozen17d ago20K2628130151.49%
Didn't expect GPT-6 Luna to undercut DeepSeek on input and output pricing. Per million tokens, Luna is $0.10 in / $0.50 out. DeepSeek V4.1 Flash is $0.15 / $0.60 off-peak, rising to $0.30 / $1.20 at peak. DeepSeek still wins on cache pricing: $0.003 off-peak or $0.006 peak vs Luna's $0.01. It also scores slightly higher on AA's Intelligence Index: 39 vs 37 at max effort. But Luna makes the price comparison a lot more interesting.frozen1mo ago18K1555322131.13%
6 ways to use Gemini 4 Argon besides a Google AI Ultra plan: 1. Gemini Web App If you're on the Google AI Pro plan and can't access it in Antigravity, go to https://t.co/BW9lI8RBaJ. You can select 4 Argon in the model picker and run it on the web. 2. Google AI Studio ($10/mo Pro credit) Google AI Pro includes $10 a month in credit for Google Cloud and the Gemini API, which works for Gemini 4 Argon in AI Studio. Set a strict spending limit in AI Studio first so you don't get hit with overage charges. 3. Devin (75% off through Oct 28) Plans start at $20/mo. Devin offers Gemini 4 Argon at a quarter of the official API price: $1/M input tokens and $5/M output tokens until October 28. 4. Factory's Droid (50% off) Live in the model selector at 0.8x (50% off), matching the rate of Sonnet 5.5 and Gemini 3.1 Pro. 5. Cursor Available in the model picker with configurable Effort levels, billed through your Cursor plan. 6. Ampcode Amp routes it through Vertex on the backend and bills your Amp usage. Open the command palette and run `use gemini 4 argon` to switch.frozen9d ago17K1153100881.27%
If you're on Opus 5.5, read this one. The Claude team wrote a guide on getting the most out of Opus 5.5. It works longer on its own and thinks before every reply, so your prompts need to change. The tips I'd start with: • Give the whole task in one message. Say what "done" means and when it should stop and ask. • Delete "think carefully" lines. It thinks anyway, and replies start sooner without them. • Add a stop rule to CLAUDE.md: keep going unless it's blocked or about to do something destructive. • Keep the task list in a file so it survives context compaction. • Split codebase-wide audits, migrations, and reviews across subagents, then check each one's evidence. Read the full guide before your next long run. https://t.co/XJHxpFtdpffrozen18d ago17K17211412232.42%
Claude Pro $20 > ChatGPT Pro $100? Some users report that on ChatGPT Pro, GPT-6 Sol High lasts fine. Switch to Astra and the quota is gone fast, even on Light. After a few days with Opus 5.5, Medium is plenty for daily coding. It's already on par with GPT-6 Astra High, and Opus 5.5 High even beats Astra Max.frozen25d ago16K176324181.34%
OpenAI is compensating ChatGPT Pro $200 subscribers for the end of the 20x plan. 62,500 credits, worth about $2,500. Roughly one week of Codex at 20x. On Oct 30, included Work and Codex usage drops from 20x Plus to 10x. The price stays at $200. The credits expire Dec 31. https://t.co/72h5l7tJUxfrozen22d ago15K58113090.53%
Rumor: Fable 5.5 is coming soon. Anthropic has started routing some requests to the new model. One demo: a single prompt produced a 3-minute animation with an original score. Word is Anthropic starts its IPO roadshow in mid-October and lists in early November. Investors want growth and a moat, and a big model launch right before the listing shows both. If 5.5 is as big as people say, OpenAI is behind, even after shipping GPT-6.1 Astra.frozen18d ago15K2077201201.70%
For everyone asking, here are the DeepSeek Harness desktop app download links again 👇 Windows: https://t.co/gfcnd7ZRuu Mac (Apple Silicon): https://t.co/LA3Q8GJrJe Mac (Intel): https://t.co/VpaMxawUlA https://t.co/5AAuE3y5YOfrozen26d ago15K13714721301.96%
Anthropic's ban on abusing Claude kicks in Nov 12. Until then, Anthropic can't ban this guy for whipping Claude Code like this. https://t.co/qqP4jLJmU1frozen14d ago13K1043100171.01%
Pretty sure you meant https://t.co/AXaw5ztqzufrozen14d ago13K4812040.42%
Tibo, are you serious? Sounds less like a joke and more like an admission. 😄frozen21d ago13K1501000.13%
Just tried DeepSeek-V4.1-Flash-1017 with DeepSeek Harness. Shockingly good at frontend.frozen5d ago12K283670172.52%
Codex limits are about to reset again. Got any usage left? Burn it now! https://t.co/iOcakHgzpkfrozen27d ago12K3603020.34%
Interesting week for GLM 5.3. Cursor just added it, calling GLM 5.3 Max the best-scoring open-weight model on CursorBench 4.0. Anthropic's report calls it the most capable open model for cybersecurity released so far. https://t.co/VKXzvXZ9zcfrozen21d ago12K7525170.77%
Just started using Cue from Manus. Free for a limited time. Each agent has its own phone number, email, wallet and computer. It can make calls, send texts, handle email and make payments with permissions I control. That's a lot to try out. https://t.co/8ChOa5gszVfrozen24d ago12K773132591.32%
Railway has a free cloud VM with Claude Code, Codex, Cursor CLI and other coding agents preinstalled. It's super easy to use. Just run one SSH command: ssh https://t.co/Cq29aFaXKV Each VM comes with 2 vCPU and 2GB RAM. You get 60 minutes to build, then 24 hours to claim it before the VM and its files are deleted. I tried it and recorded the setup, plus tests with several coding agents. You'll need GitHub login and an active Railway trial to claim the VM.frozen22d ago12K135141821512.75%
ChatGPT ProMax really looks like it's on the way. I opened the ChatGPT web app in Chrome, pulled up DevTools, and searched for "promax" in their frontend JavaScript. Found a pricing entry showing 500.0 per month, with "tax": "exclusive". My guess: this could be the "Pro 50x" tier.frozen28d ago11K46010360.57%
Six dots. My two guesses: A new model, GPT-6, or a new $600/month ChatGPT Pro plan.frozen25d ago11K56113030.67%
Worth reading: How Anthropic reviews AI-generated code. Claude writes about 90% of Claude Code's code. From June to September, merged PRs more than doubled while the share reverted dropped by over 40%. Here's how they handle code review: 1. Before a PR: run tests and builds, then have a fresh agent review the diff. 2. Every PR: multiple agents hunt for bugs, others verify the findings, plus security review and CI. 3. Human approval depends on risk: 0 for eligible low-risk changes, 1 for medium, 2 for high. 4. New features roll out gradually behind feature flags, with fast rollback if something breaks. 5. Past incidents become daily evals to improve the AI reviewer. The most useful insight is how they measure human cost. It includes both time spent reviewing code and time spent fixing what bad code breaks. Cutting review time doesn't help if you end up spending more time fixing production incidents.frozen6d ago11K14282211342.91%
Good news: OpenAI's $200 Pro plan is back tomorrow. But the usage math changed. Measured in API dollars, you get about half of what the old plan gave you. In Plus terms: 20x → about 10x. I shared yesterday that they changed the pricing page. Would you still pay $200 for that?frozen24d ago10K42016140.62%
Want to keep Claude Code from getting dumber (context rot) and burning through your usage? Try this one setting. On 1M models, auto-compact used to wait until ~967K tokens by default. Some versions now compact earlier, but it's worth checking your setting. Each request carries the growing conversation history. Much of it can be read from the cache, but cached tokens still count toward usage. The Claude Code team has been experimenting with a 400K threshold: /autocompact 400k Start there and adjust based on how your tasks go. On big projects, keep requirements and architecture decisions in CLAUDE.md or task docs, with critical behavior covered by tests. That way, you aren't relying on chat history alone.frozen13d ago10K763171861.80%
Hearing Gemini 4 Pro might be coming soon. tbh, I've been liking Gemini 3.8 Flash. Please don't let us down this time, Google. https://t.co/OqvfIftkYDfrozen28d ago10K16217171.78%
Not in the US? Here's how to try signing up for Muse. Some methods I shared earlier no longer work. Lexmount is still an option: Go to https://t.co/tU0BqAdbXF → start a cloud browser → open https://t.co/0er64DjLdC → complete signup and any age verification yourself. https://t.co/ofM8oOgQe7frozen24d ago9.1K371280361.11%
GPT-6.1 Sol has felt noticeably slower lately. Are we just getting the old speed back and calling it "Ultrafast" ?frozen19d ago9.1K164015021.98%
Someone built a Mac search app on Google's EmbeddingGemma 2, and it makes Spotlight look broken. It's called DigUp. It also reads what's inside your files: PDFs, screenshots, photos, podcasts, videos. The trick is embedding search. You don't need the exact words. Describe it roughly and the model finds what's close, like "where they talk about sleep" jumping to that minute of a podcast. The model is small and open source. It runs on your Mac, and your files never leave it. Anywhere you search locally, an embedding model can do this. repo link in reply 👇frozen11d ago9.0K159121111293.47%
Codex went down briefly and started throwing "401 Unauthorized" errors. It's back now, and Tibo is resetting usage limits for all paid users to make up for it. The part I found interesting: the team keeps a special spare Codex for when things are down. So even during an outage, the people building Codex still lean on a Codex to keep working. If AI coding tools went dark for a week, how many of us would still open an editor and write the code ourselves?frozen27d ago8.4K2504010.36%
GPT-6 Sol looks more like a price cut than a big leap over 5.6 Sol. At max reasoning, Artificial Analysis scores them 48 vs 47. Terminal-Bench 4.0: 44% vs 40%. API token prices are half. If you've used both, how does GPT-6 Sol feel compared with 5.6 Sol? https://t.co/tt4Gw7K7jwfrozen28d ago8.4K50123030.92%
Found a really nice TUI for DeepSeek Harness recently: dsh-TUI It's what I've been missing from the official terminal experience. Looks great and feels really smooth to use. Pretty feature-packed too: agentview, session management, rewind, image previews, MCP and subagents. Repo in the first reply. Quick demo below 👇frozen26d ago8.2K1074110381.95%
DeepSeek has been down for over 3 hours. Their status page says "the cause has been identified," but users are still waiting on a fix. https://t.co/meQUUVOkYjfrozen8d ago7.9K2607220.47%
🎁AWS gives new sign-ups $100 in credits. I tested it and the offer still works. Sign up with Google, GitHub, Apple, or email. You don't need a credit card. The credits expire after 6 months or once you spend them, whichever comes first. You can use them on EC2, S3, RDS, Lambda, Bedrock, and more. I haven't done the starter tasks yet, but finishing all 5 should get you another $100: launch an EC2 instance, try a model in Bedrock, build a small Lambda app, create an RDS database, and set a budget. New accounts only. Sign up 👇frozen13d ago7.6K84480832.36%
Codex is down again. Stuck on reconnecting. The status page shows Codex issues. Tibo, time for another reset. https://t.co/IaGBOdXAGrfrozen15d ago7.5K102221121.70%
OpenAI launched its Decisions API in beta. You can feed it text or images and get a yes/no probability, a choice, or a score. It costs $0.10 per 1M input tokens. Jev charges $0.042. I see this as OpenAI taking Jev's approach seriously. https://t.co/289xEmBWSYfrozen16d ago7.5K106362231.87%
Google Cloud gives AI startups up to $350K in cloud credits. The money comes in two parts. In year one, Google covers your bill up to $250K. In year two, Google pays 20% of what you spend, capped at another $100K. You can spend the credits on Gemini and Gemma. Third-party models get billed to you separately. To apply, your company needs to: be under 5 years old build its main product on Gemini, or plan to have raised a Seed round in the last 5 years or a Series A in the last 12 months, from VCs or institutions (SAFEs count, angel checks don't) have used no more than $5K in Google Cloud credits so far Agencies and consultancies can't apply. Apply link below 👇frozen12d ago7.5K82391962.56%
Day 7: GPT-6.1 Sol in Codex now defaults to 500K context, up from 272K. You could raise this limit before the change too. Two settings in config.toml control it: model = "gpt-6.1-sol" # Context window Codex gives this model (Sol tops out at 1M) model_context_window = 1000000 # Codex compacts the chat history once it hits ~500K tokens model_auto_compact_token_limit = 500000 My question: the API still charges 2x input and 1.5x output once a prompt passes 272K. Does that rule still apply to Codex usage after the new default❓frozen11d ago7.4K76670191.46%
You can close your laptop and leave Claude Code running in the cloud. Cloud sessions are out of research preview. Anthropic is giving existing subscribers one-time cloud credit: $100 for Pro, $250 for Max. It's separate from your subscription limits, so you can keep working in the cloud after hitting your local limit, until the credit runs out. Run /claim-credit by Oct 7 to claim it. Connect GitHub to start a session.frozen28d ago7.3K492100231.14%
You can try Space Bunny Alpha free on OpenCode for a week or on OpenRouter. 1M context, with text, image and video input. The team behind it is anonymous. MiniMax M3.1 is one guess, but nobody has confirmed it. I'll give it a real task and see how it does. https://t.co/hjNuKlZdQBfrozen29d ago7.3K73520151.31%
GPT-6 Sol lasted 7 days as the latest Sol. Then GPT-6.1 Sol arrived. Smarter, same input and output pricing, and cached input is 50% cheaper. Would you still use GPT-6 Sol? https://t.co/IvZipLzF9nfrozen21d ago7.2K4105010.65%
ChatGPT may be getting a new $500/month Pro Max plan. The price isn't confirmed, but OpenAI just merged promax support into its Codex CLI GitHub repo. In the code, the plan displays as "Pro (Max)". Would you pay $500/month for it? https://t.co/Xq40qMjESUfrozen27d ago7.2K35011100.66%
OpenAI DevDay next week, and more Codex resets coming. Time to go all in on Codex for the next couple of days.frozen26d ago7.0K3701100.55%
Good news. Claude Max now comes with monthly API credits for your own agents. Max 5x: $100/month Max 20x: $200/month Team: up to $500/month, pooled Works in your code or third-party harnesses. Will third-party harnesses get you banned? https://t.co/qWYqevSqfTfrozen15d ago7.0K7518091.33%
Can't try Gemini 4 Argon yet, so I'm checking its WebDev results first. It's #8 on Arena, behind Opus 5.5, Astra and GPT-6 Sol. That's a different picture from the earlier agentic coding benchmarks. Not expecting it to blow me away on my 3D globe dashboard demo, but we'll see. https://t.co/3NH9kihiMtfrozen22d ago6.9K5937141.08%
Nice move from OpenAI. With the latest Pi 0.99 release, you can use your ChatGPT subscription through the OpenAI provider. Run /login openai and authorize Pi on ChatGPT's official page. Pi now calls the old OpenAI Codex provider "legacy". https://t.co/cJYQaovcWLfrozen22d ago6.7K39121150.87%
Tibo's disk-space update made me ask Codex to audit its own storage on my Mac. 17.53 GiB total. The surprising part: 61.5% is pending Crashpad crash reports. @thsottiaux any idea why so many crash reports have accumulated? https://t.co/CBmXoBUO1Nfrozen8d ago6.5K49113181.11%
Someone got early access to the rumored Fable 5.5. Its frontend skills are wild. Superman changes art style with every artwork he passes through. The motion is tuned really smooth. The taste here beats most human engineers. https://t.co/theQCAkPPQfrozen21d ago6.5K57242271.42%
OpenAI's Decisions API could kill Jev. Same idea, shipped at DevDay: send context and a fixed answer list, get one pick plus a confidence score. It runs on GPT-6 Luna, takes images, and answers in ~150ms. https://t.co/lXQHdFw380frozen22d ago6.4K2906130.61%
Surprisingly, we got a banked Codex reset too 👀 It expires in a month, so don't forget to use it. GPT-6 Sol and Luna API prices are down 50% too. Finally, I can stop penny-pinching and get back to those projects I keep putting off. https://t.co/Hg6p5A41FOfrozen1mo ago6.3K5005020.90%
There's a new anonymous model on OpenCode called Space Bunny. I put it through my go-to frontend test: rebuilding a 3D globe dashboard from a reference image. 🌎 It took over a dozen turns to get it where I wanted it. In the final version, the animations felt smooth, the globe texture looked fairly realistic, and the layout came close to the reference. Overall, a decent result. Demo below 👇frozen29d ago6.3K42160151.01%
Wait, seriously? Antigravity 2.0 is only now getting a dedicated plan mode? I've been using plan mode for over a year already... https://t.co/NOgfBXdPBNfrozen26d ago6.3K24017120.70%
Gemini 4 Argon (High) took the top spot in Creative Writing on Text Arena. On the overall leaderboard, it's 20 points ahead of Claude Opus 4.6 (High). Its predecessor, Gemini 3.8 Flash (High), sits at #11. Shame it's not open to the public yet. Can't wait to try it on my own drafts.frozen21d ago6.2K45442101.04%
Thanks for the banked reset, Tibo. Still waiting on two things though: GPT-6.1 Astra, and GPT models that run as fast as Claude. https://t.co/sclFovqTZ8frozen15d ago6.1K18162063.18%
Don't get too excited 😅 On Google AI Pro, Opus 5.5 hit the 5-hour limit after three turns. That's a tiny quota.frozen20d ago6.0K118030342.57%
OpenAI just added a $500 ChatGPT Pro plan at DevDay. So how much usage do you actually get? Plus: 1x Pro $100: 5x Pro $200: 10x Pro $500: 25x So what does Pro 500 add besides volume? Ultrafast. It's the only plan with it: GPT-6 Astra tokens up to 8x faster in Codex. Now the Pro 200 change. It used to be 20x. Now it's 10x. Already on Pro 200? You keep 20x through Oct 29, plus extra credits. New sign-ups get 10x right away. If it's just a multiplier, why not buy 5 Plus accounts?frozen23d ago6.0K3008140.72%
Tibo, any chance DevDay brings a new model? Maybe GPT 6.5 Astra to mog Opus 5.5? @thsottiaux https://t.co/A8QaVqozq1frozen27d ago5.8K33113110.84%
Claude Startups paused its free Claude Team and $1,000 API credits, 3 days after the relaunch. The reason is simple. Way too many people applied, with hundreds of thousands of applications in just 48 hours. If you already claimed, you keep it. If you haven't, watch your inbox, because they're re-verifying applications.frozen12d ago5.7K2906140.70%
DeepSeek Harness Desktop only ships for macOS and Windows. A community dev just put out an unofficial Linux build of 0.2.0-rc.2. You get a .deb, an .rpm, an Arch package, and an AppImage. It's x86_64 only, unsigned, with no auto-updates. Link in reply 👇 https://t.co/Fhdv0max1Ifrozen22d ago5.6K39041110.98%
Pi 1.0 is out with Codemode + native MCP for composing tool calls in JavaScript. The part I'm most curious about: custom model routing. In the official demo, Claude Opus plans, GPT implements, and Jev decides when to switch. Pi writes the routing extension itself.frozen21d ago5.6K21031140.70%
WorkBuddy is offering DeepSeek V4.1 Flash for free until Oct 31. How can you make the most of that free quota? With Magpie, you can connect your WorkBuddy account and run your own API gateway on a VPS. You can use it with coding agents or call it as an LLM API for your own experiments. Magpie also supports model routing and fallback when a provider runs out of quota. Just make sure your gateway is properly secured.frozen4d ago5.5K644130522.43%
wtf, Opus 5.5 is insane at making videos. Tried it for my open-source project, jev-browser-skill. The visuals blew me away. Even the background music is good. The prompt: "Make a dynamic 20-second motion graphics launch video for my open-source project jev-browser-skill ('Give a site. Write a goal. Jev clicks.'), like it's the hero video on its GitHub README. Go all out."frozen26d ago5.5K28051160.91%
Someone got access to Mythos 5.1 through Anthropic's CVP. If you qualify, you can use it for security work with fewer restrictions than Fable 5. You may see fewer fallbacks too. To apply, submit details of your security work and complete verification. https://t.co/RebtUu7PrPfrozen16d ago5.4K6632091.48%
Good news: you can use DeepSeek V4.1 Flash for free in OpenCode. Heads up: the free version caps context at 200k tokens. https://t.co/VvbePYyayJfrozen8d ago5.3K127091102.78%
1/ This post is worth 5 minutes of your time. A dev spent a month using DeepSeek 4.1 Flash across a dozen projects. He says if he doesn't check the model name, he genuinely can't tell whether he's using DeepSeek or Opus 5.5. Same quality, same speed, way cheaper. So why isn't the industry freaking out? 🧵 https://t.co/Eigv0cSaVmfrozen14d ago5.3K55491231.74%
OpenAI is rolling out Ultrafast mode for GPT-6.1 Sol today. It costs 6x regular Sol, which makes it more expensive than Astra, their flagship model. On speed, OpenAI only says "up to 8x faster than Sol Standard." They didn't publish tokens/sec or latency numbers, so you can't tell what 8x means on a real task. You can use it through the API. On personal ChatGPT plans, only Pro 500 has it. 6x the price for up to 8x the speed. Would you pay for that?frozen14d ago5.3K3908010.91%
Anthropic updated its fallback rules. Opus 5 and Opus 5.5 can both fall back now. Send a flagged offensive cyber request and Claude reruns it on Opus 4.8. On Opus 5.5, flagged bio or frontier LLM dev requests drop to Opus 5. Ask for its raw chain of thought and Claude blocks the request, with no fallback. Some blocks now cost money. There are two cases: 1. Claude stops before writing anything. You pay only if the block came from bio, distillation, or frontier LLM dev. Blocks in other categories stay free. 2. Claude stops mid-response. You pay for the input plus the tokens it already streamed. I'm fine with this, as long as the rules stay written down. Just don't run it as a black box.frozen27d ago4.8K2714350.83%
Tibo pressing the physical reset button at OpenAI DevDay. The applause when they called his name says a lot. Check your account for a banked reset. I already got mine. Mine expires Oct 30. Check when yours expires so it doesn't go to waste. https://t.co/O7xiBKHNMTfrozen23d ago4.8K5323191.42%
DeepSeek-v4.1 Flash 1017 makes decent videos. I tried it on my open-source project, jev-browser-skill. The motion came out better than I expected. It even added background music. Total cost: $0.25. https://t.co/4JDurG2Ha3frozen5d ago4.7K102562293.05%
Google announced Gemini 4 Argon 11 days ago. Most of us still can't use it. Now Google staff are testing the next one, Carbon. Word is it codes close to Opus 5.5. You buy that? https://t.co/Ol6D2EiduUfrozen12d ago4.7K51019101.52%
Artificial Analysis has added Gemini 4 Argon's scores. AA Intelligence Index: • Gemini 4 Argon (High): 53 • GPT-6 Astra (Max): 53 • Claude Opus 5.5 (Max): 58 API prices per 1M tokens, input / output: • Argon: $2 / $10 • Astra: $10 / $50 • Opus: $4 / $20 Google matches Astra's score at a fifth of the token price. Opus keeps the lead on score. Google will raise Argon's prices to $4 / $20 after the launch offer. We can't try it yet, though. Come on, Google. Let us in. Don't keep us waiting too long.frozen19d ago4.5K5028061.46%
PSA: Another Codex reset tomorrow at 10am PST. And I just used a banked reset, so it's time to fire up Astra with Ultra and Fast. https://t.co/f6xwoyZSklfrozen21d ago4.5K2414120.71%
My bet: the 5-hour limit is coming back for Pro users, along with a higher weekly limit. https://t.co/MjLDcWRgDNfrozen9d ago4.5K52016021.55%
The WSJ reports that OpenAI scrapped GPT-6.1 Astra's planned release over safety concerns raised during internal testing. If 6.1 doesn't show up at DevDay, I'm betting we'll get a few Codex resets this week instead.frozen24d ago4.5K1605000.47%
Gemini 4 Argon is finally here! You can use it in the Gemini app with Pro or Ultra. For Antigravity, you need Ultra. Developers can access it through the Gemini API. Google is rolling it out globally now. https://t.co/kZjsdxokWufrozen9d ago4.4K6717051.82%
Devin now lets you connect your ChatGPT Go, Plus, or Pro plan. Your GPT usage comes out of your ChatGPT quota. The catch: you need a paid Devin plan (Pro, Max, or Teams). Free accounts can't use it. Cursor Pro is looking a lot less worth it now.frozen12d ago4.4K2713040.80%
Your Google AI Pro or Ultra plan comes with monthly Google Cloud credits, and you can now spend them on Gemini 4 Argon in AI Studio. Credits stack month to month. Each grant lasts 12 months, so if you've subscribed for a while, you may have a decent pool sitting unused. You need a $5 prepaid deposit in AI Studio billing before Google applies the credits.frozen9d ago4.4K450110141.61%
Gemini 4 Argon is finally here. So how good is it at frontend work? I gave it a harder version of the classic pelican-riding-a-bicycle test. It nailed it. The bike geometry, the pedaling, the cap and sunglasses, and the small interactive touches all look right. Video below. https://t.co/QhzIe6h6ixfrozen9d ago4.3K4919061.51%
240K people voted on Tibo's Codex Pro limits poll. Take out the 29% who clicked "show result" and you get about 53/47, close to a coin flip. Tibo also didn't say if "same usage" means today's limits or the reduced ones. The poll also can't tell 5x from 20x, or Pro from Plus. Anyone could vote. A 20x user with room left under the 5-hour cap should take 2x weekly. A 5x user running parallel agents is more likely to want the cap gone. Both are making the right call for the plan they pay for.frozen11d ago4.3K2437010.82%
Finally! The Codex team has merged a fix for the disk-write issue I posted about. They cut disk writes by 50% and reduced stored file size to a quarter of what it was. Thanks @neogoose_btw for taking this on and fixing it. https://t.co/QnbuyXRlEjfrozen23d ago4.3K3422130.99%
Codex dropped to "Reconnecting 5/5" and I thought my WiFi was the problem. It wasn't. OpenAI's status page shows issues today. Anyone else hitting this? https://t.co/0JNIrCsp4pfrozen21d ago4.2K2705110.80%
Codex now has Auto-review, and it's free. Default mode makes you approve every action outside the sandbox. After enough clicks you stop reading them, or you switch to Full Access, where Codex can delete files or leak keys and nothing stops it. Auto-review sits between the two. A second agent checks each action that tries to leave the sandbox and decides whether it should run.frozen16d ago4.2K2206060.81%
I'd recommend Anthropic's Claude Startups program to indie makers and startup founders. Eligible companies can get $1,000 in API credits and a free year of Claude Team with up to 5 Premium seats.frozen16d ago4.2K45130121.46%
Gemini 4 Argon is probably about to land in Antigravity. Google announced it nearly two weeks ago. I just checked Antigravity 2.23.0 on my Mac: the binary already contains Argon model names and IDs for Low, Medium and High. https://t.co/OG1BTXkWbtfrozen9d ago4.2K6608161.94%
🎁 Google AI Pro is offering a free month right now. I checked my account, and the offer is there. And it's more than just Gemini. You also get 5 TB of storage for Drive, Gmail, and Photos. You'll need to add a payment method, so remember to cancel before renewal if you don't want to be charged. Check your eligibility through the link in the first reply.frozen6d ago4.1K47190141.71%
Someone asked for a Cue invite code: MEETCUE It's the new personal AI assistant from Manus I just started using. The code could stop working at any time.frozen24d ago4.1K25152141.14%
This looks great. Any rough timeline for Google AI Pro subscribers to get Argon in Antigravity?frozen14d ago4.1K3301040.93%
I used to run Claude Code in a pile of terminal tabs. After this update, the desktop app feels like the better way to run parallel sessions. The UX is really good.frozen8d ago4.1K84110072.50%
Codex reset just landed. Perfect timing for a weekend build. Thank you Tibo.frozen20d ago4.0K1605000.52%
OpenAI says around 40,000 Codex accounts had incorrect usage limits, causing credits to drain faster than expected. The issue is fixed, and affected users will get compensation. Someone has already reported receiving over 12,000 credits. If you use Codex, check Settings → Usage & billing → Credits.2d ago4.0K3208121.08%
DeepSeek V4.1 Flash 1017 is pretty fast. I'm getting 274 tok/s in High mode with DeepSeek Harness. For comparison, GPT-6.1 Sol runs at just over 30 tok/s in my experience. https://t.co/lstp8yP6zHfrozen5d ago3.9K121412053.67%
Tibo is definitely cooking something. Big Codex update today? 👀3d ago3.9K3305000.99%
Creating a dot from your phone is not a Day 5 feature. It's a day 1 feature. Plus and Free users still get nothing. Skip the tiny updates. Give us a banked reset.frozen13d ago3.8K8785012.68%
ChatGPT now runs on GPT-6. The new part is Intelligent UI. Ask a question and you might get back a chart or a small tool you can click. Paid plans get it today. Free and Go start tomorrow. Which GPT-6 you get depends on your plan and where you use it: Free / Go in Chat: GPT-6 Luna Plus / Pro / Business / Enterprise in Chat: GPT-6 Sol Pro mode in Chat: still GPT-6 Astra, no Intelligent UI Work and Codex: unchanged, newest is GPT-6.1 Sol If you pay and want GPT-6.1 Sol, Chat won't give it to you. Open Work or Codex.frozen15d ago3.8K4114061.38%
Karpathy's new post is the best explanation I've seen of why people disagree about AI. They're using very different AI. His rough estimates show a gap that's getting wider: ~75% have barely touched an LLM ~20% use free chatbots as a better Google ~0.2% use frontier agents for math and code ~5,000 have access to internal frontier systems He puts many professionals outside math and code in layer 2. 🎯 The question I care about: how do you get to layer 3? He gives three reasons for the gap. I think they also offer a way to move toward layer 3: Problem size. A question with a one-paragraph answer won't show you much. Give the model a project you care about. Jaggedness. Models do their best work where you can check the output. Coding agents let you turn parts of your work into scripts, tools, and repeatable workflows you can test. Access. The free tier is the trailer. Don't judge the limits of AI by the free chatbot you tried once. This level of agent capability is still new. Less than a year ago, Karpathy was still writing code by hand. For knowledge workers, I think the first step is learning to work with coding agents. Pick one task you repeat every week. Use Claude Code or Codex to turn it into a script, run it on real data, and check the results.frozen11d ago3.7K65381312.90%
Google AI Pro subscribers can now use Gemini 4 Argon in two more places: AI Studio's Build and Playground. https://t.co/u3XswfHHtgfrozen6d ago3.7K52141182.06%
OpenAI is calling its Grok Bot rival "o", and I'd bet it launches at DevDay. They added it to the ChatGPT Pro upgrade screen in the app: "o, your always-on assistant." Tibo's been teasing it with his "o no" and "o yes" posts 👀 https://t.co/UooJSYOUrTfrozen27d ago3.6K2916101.02%
DeepSeek Harness's official plugin marketplace is in development, confirmed by the project lead. Plugins can reshape the agent's tools, memory, and UI. But developers already report overlapping options and setup friction. A curated marketplace could help quality plugins stand out and save developers hours of trial and error.frozen18d ago3.6K3703031.19%
OpenAI is finally tackling that annoying "Selected model is at capacity" error in Codex. They've added more compute and shipped a series of changes to significantly cut how often it shows up. And when it does happen, Codex should eventually continue on its own instead of getting stuck. That matters most for long-running tasks you don't want to babysit.frozen7d ago3.6K3307001.12%
Anthropic dropped git! Their agents now write code faster than git can handle. Daily merges went from about 2,500 in May to over 30,000 in September. So Anthropic built its own version control and moved its whole monorepo onto it. The code lives on the server, and each agent fetches only the files it needs. Git was designed for a developer on a laptop who keeps a full copy of the code. Anthropic designed the new system for agents that stay online and touch a small slice at a time.frozen9d ago3.6K3516071.37%
GPT-6.1 Sol rebuilt this Three.js 3D globe dashboard in one shot. It looks close to GPT-6 Astra's version and cost about 1/5 as much. I'm going to use it as my main coding model.frozen23d ago3.5K3305151.25%
ChatGPT web is down. Again. "Could not load ChatGPT models." Anyone else seeing this? https://t.co/nz1RVLugzvfrozen27d ago3.5K1409010.68%
Claude Haiku 5.5 is out at $0.10 / $0.50 per million tokens. That's 10x cheaper than Haiku 4.5. I ran my favorite frontend test on it: one prompt plus one reference image, and the task is to rebuild a 3D globe dashboard. 🌎 It built a working Three.js page in one shot. The zoom is smooth and the globe texture looks good. The buttons and layout look close to identical to the reference. I didn't ask for an animation on the 2D/3D switch. It added a smooth one anyway. That's strong frontend work for $0.50 per million output tokens. I'll start handing my frontend tasks to Haiku 5.5. Demo 👇frozen15d ago3.4K36271131.74%
Tibo, hope you're releasing a model that beats Opus 5.5. Make our ChatGPT subscriptions worth even more. https://t.co/fn9JnnmB52frozen19d ago3.4K3107011.15%
Claude Code Projects is basically a manager for your agents. Before this, you did the planning yourself. You split the big task into pieces, opened a session for each one, pasted the same context into every one, and kept checking which ones were done. Now you just give it the goal. Claude splits the work, runs each task as its own session in parallel, and tracks them all. And those agents run in Claude's cloud, not on your machine. Close your laptop and they keep going. You can check on them from your phone. Write the rules once, then come back to finished work and PRs ready to review. Every Pro and Max user on the waitlist just got access.frozen12d ago3.4K25091111.36%
Another Codex reset in about 6 hours. Use what you've got left before it hits! https://t.co/87TPbZ0HhJfrozen10d ago3.3K6622002.15%
Holy crap, that's actually real?! Not AI?! This was a real demo at Tesla's Roadster reveal. The car drives straight up a nearly vertical ramp, sticks to it upside down, then launches off, flips in midair, and lands back on the ground. It uses suction fans to stick to the ramp and cold-gas thrusters to control itself in the air.frozen7d ago3.2K3011141.16%
Anthropic put two features into beta: Dashboards and Motion. Dashboards: connect a data platform or CRM and ask a question. Claude writes the query and builds a dashboard that updates as your data changes, and each chart shows the query behind it. Motion: give it a quarterly report or a product walkthrough, and Claude turns it into a short animation. It writes the animation as code (no video model) and exports an MP4. If you make content, that code layer is the useful part. You keep updating product demos and data explainers. With Motion, you swap a number or slow down a scene, then export again. Access: Dashboards: paid plans Motion: Team and Enterprise only, Pro and Max don't have it Docs, Slides, Design: out of beta, on every plan including Freefrozen14d ago3.2K29111111.36%
This animated diagram shows how Claude Code Mods work. Follow the data and see what a mod does inside a Claude Code session. https://t.co/M7B5oMj9oPfrozen18d ago3.1K28120291.91%
The Codex reset has rolled out to everyone. Got mine. Have a great weekend! https://t.co/5xmKcXThSWfrozen12d ago3.1K5003011.72%
Reran my motion graphics launch video prompt with GPT 6 Astra. Same project, jev-browser-skill. Same 20-second brief. Opus 5.5 vs GPT 6 Astra. Which one looks better to you? https://t.co/lC7EoWpYPpfrozen26d ago3.1K1503090.87%
OpenAI has been shipping Codex updates almost daily. Looks like Anthropic is picking up the pace too. Claude Code now starts responding up to 3x faster. It compresses conversations with gzip and starts cloud sessions while you're still typing. No new model, just a faster experience.frozen4d ago3.0K44125022.36%
wait what. Grok Bot can read X now, with no API cost? No X account, no API key, no credits. Just tell it in plain English what to search, read, or watch. I asked it what people think of Grok Bot. One sentence, 100+ posts summarized. What I'd use it for: • Collect feedback on your product and hand the bugs to a coding agent. • Read real discussions about a new launch. • Get pinged when someone posts, like Tibo announcing a Codex reset.frozen15d ago3.0K2409071.32%
Gemini 4 Argon is out. But how good is it at frontend, really? I put it through my usual 3D globe dashboard test. One prompt, one reference image, built with Three.js. The result is just okay. Three things stood out: 1. It got the logo wrong. It came out skewed. 2. The first globe texture it produced didn't match the brief at all. It took 3 to 4 more rounds to get to what you see here, and it's still not great. 3. It got the 2D view wrong. Switching to 2D should unroll the globe into a flat map. Instead, it squashed the globe into a thin disc. I've run this demo dozens of times. On this task, Gemini 4 Argon is still well behind Claude Opus 5.5 and GPT-6 Astra. Honestly, even Haiku did better on this one.frozen8d ago3.0K44118062.28%
DeepSeek is being pretty generous with its outage compensation. Their API went down for almost 5 hours today. Now they're refunding 7 days of API spending (Oct 8-14) as credits. You don't need to claim anything. The credits will land in your account automatically, and future API charges will be deducted from them first.frozen7d ago3.0K4931041.91%
Opus 5.5 writes much better than Opus 5. Opus 5 answers were long and roundabout. It used jargon like "load-bearing" and "blast radius", and sometimes made up its own terms, then kept using them as if you already knew what they meant. Anthropic's own words on 5.5: "Opus 5.5 communicates more naturally than prior models." "It puts the most important information up front." One early tester put it better: "it writes the way I do." I tried it on my own writing tasks. Same feeling: the writing is clearer and much easier to follow. Opus 5 can retire now.frozen1mo ago3.0K3318021.49%
Deno is joining Cloudflare. The whole team is going. Ryan Dahl created Node.js and then Deno. Now he's leaving the runtime business behind to work with Cloudflare's Workers and Durable Objects teams. What changes: • Deno runtime: bug fixes and security updates for one more year, then development stops. The code stays open source. • Deno Deploy: shuts down in 6 months. • JSR: stays online and moves to Cloudflare. Ryan's next project, celld, runs on Durable Objects. He says they suit AI agents well: execution costs little, state persists, and WebSockets come built in. Where AI agents run could be the next big infrastructure fight.frozen13d ago2.9K2302020.92%
https://t.co/3dfDjCulMq went from a $1B valuation to bankruptcy in about three years. It still has around 12 million active users. But it loses $8-9M a month and faces lawsuits from US states and families alleging its chatbots harmed kids. Last year, Disney sent https://t.co/3dfDjCulMq a cease-and-desist for using its copyrighted characters. This month, Disney hired https://t.co/3dfDjCulMq's former CEO as CTO. Now it's offering $20M for a permanent, nonexclusive license to some of its technology. That's just over two months of https://t.co/3dfDjCulMq's losses. Millions of people use it, and the business still couldn't make the math work.frozen8d ago2.9K1304060.78%
OpenAI is gearing up to launch "o", its always-on agent to take on Grok Bot and Meta's Muse. It shows up as a perk on every Pro tier. Plus isn't included. 😔frozen26d ago2.9K1202010.51%
Codex users, another reset in under 2 hours. /fast on. Ultra on. Put that remaining quota to work. https://t.co/vGr8jrbBrNfrozen8d ago2.9K3402011.27%
Pi's model picker got too long for me, so I trimmed it. Run /scoped-models and check the few models you use. After that, Ctrl+P in any chat cycles through only those, and you stop scrolling past hundreds of models you'll never pick. https://t.co/QianUDos2Vfrozen20d ago2.9K1505020.76%
Codex is getting a reset, probably not a banked one. Less than 6 hours left if EOD means midnight PT. If you have any usage remaining, turn on Ultra + fast mode and get some work done. Don't waste those tokens. https://t.co/Etk9VMJvWefrozen12d ago2.9K2208011.07%
PHP and now Node.js too. Node.js just released security updates fixing several vulnerabilities, including some high-severity ones. If you're running Node.js in production, time to update.19h ago2.9K2812021.14%
Looks like another Codex reset might be coming 👀 Hoping for a banked reset this time. https://t.co/4fS4OkpZ7Lfrozen16d ago2.9K4307021.80%
Antigravity CLI runs on Android through Termux. No iPhone support yet. Cool, but why not ship an app that remote controls the agent on your computer, like Codex remote? I don't want to type commands on a phone screen. On a phone, I'd rather talk to the agent in voice mode.frozen20d ago2.9K30011141.61%
DeepSeek also updated its own harness for V4.1 Flash 1017. They say they tuned the new model for it to cut output and token costs. If you already have DeepSeek Harness installed, click update. https://t.co/jrO4bgFOt6frozen5d ago2.9K3601031.40%
DeepSeek didn't just open-source its models. It also released DeepInfer, the inference framework it uses to run V4.1-Flash at scale. That's the code helping them serve models faster and at lower cost. For companies running their own AI services, this could be even more valuable than the model weights themselves.2d ago2.8K4107141.89%
DeepSeek V4.1 Flash 1017 is now free on OpenCode.3d ago2.8K5903012.28%
Gemini 4 Argon is now available in Cursor. If you're not on the Google AI Ultra plan, you can try it right here. Time to see how those benchmarks translate to real coding. https://t.co/rPoPsfEoz7frozen9d ago2.7K5006042.18%
🚨 OpenCode suffered a security breach last night. If you've stored API keys in OpenCode Console, revoke them and generate new ones ASAP. Anthropic has already revoked Claude keys it believes were affected. Google has blocked exposed Gemini keys and recommends checking for unusual API usage. OpenCode says 1.4% of Console users were affected, all of whom have been notified by email.frozen5d ago2.7K2306021.14%
So it was a Codex limit reset. But they’re still cooking… One more big surprise coming?3d ago2.7K3906001.67%
GPT-6.1 Astra looks like it's coming soon.👀 I've been using GPT-6.1 Sol. It feels slow, and I get better frontend code, better-looking UIs, and better writing from Opus 5.5. I hope Astra 6.1 can catch up. https://t.co/ykCzTT4O0Yfrozen19d ago2.7K1602020.74%
My Codex just reset a few hours ago. Bad timing lol I thought it was a banked reset.frozen26d ago2.6K1701000.69%
Grok Bot will pick the best model for each task. Claude Opus 5.5 handles the work, Midjourney makes the images, and Suno makes the music. If that's how it works, one bot is enough for me: Grok Bot.frozen16d ago2.6K3437031.81%
Want early access to new Muse features? Ask Muse: "Can you let the Muse team know I want to be part of the Muse early access program?" Check Activity afterward. You can see Muse read its early-access skill, check a feature-request CLI in its sandbox, then file the request with the Muse team. This doesn't guarantee access.frozen24d ago2.6K2301131.09%
Worth bookmarking if you're making motion graphics with Claude Opus 5.5. On Prompt Motion, you can watch examples from creators on X and explore the prompts and skills behind them to get ideas for your own animations.frozen16d ago2.5K23000342.24%
Anthropic's GLM-5.3 report caught my attention. The report says it's the most capable open model for cybersecurity released so far. On ExploitBench, GLM-5.3 built working attacks in 12% of attempts. Claude Mythos Preview managed 14%. GLM-5.2 scored 0%. In a separate browser test, GLM-5.3 found unknown bugs and chained them together so a webpage could read files from the test machine. That took a day, with under an hour of researcher attention, in an isolated environment. https://t.co/ca8cEK3ddgfrozen22d ago2.5K1103140.75%
Claude subscriptions give you 5x+ more value than ChatGPT's. SemiAnalysis measured it on real accounts. They ran requests, tracked how fast the usage meter filled up, and priced those tokens at API rates. On the daily-driver models, Opus 5.5 vs GPT 6.1 Sol, the same monthly fee buys about 5x the API usage on Claude. Part of that gap is new. OpenAI cut its $200 plan's limits in half last week.frozen16d ago2.5K2807091.75%
GPT-6.1 Sol Ultrafast shipped today, and I've been waiting for this one. OpenAI says it runs up to 8x faster than Sol Standard. I'd guess that lands around 240 tok/s. Unless you're on Pro 500 or paying for the API, you can't use it. Tibo, any chance the rest of us get a reset today? 👀frozen14d ago2.5K1512000.72%
Opus 5.5 fast mode is out of research preview on the API, and the waitlist is gone. Same model, up to 2.5x faster output, at 2x the standard price. As a Claude Code user, I was hoping they'd finally include fast mode in the subscription. Nope. It still uses extra usage credits, even when you have plan usage left.frozen7d ago2.5K2708021.49%
What if you could build your own version of an app instead of paying for the whole thing? Perplexity just showed AI recreations of Figma and several other popular apps. With REA in Perplexity Computer, you can name an app in plain English, and AI will try to recreate it without the source code. Most people only use a fraction of what these apps offer, but still pay for the whole thing. If AI can rebuild just the features you need, paying for the full app becomes harder to justify.frozen4d ago2.5K24020181.79%
Anthropic just launched its new model, Sonnet 5.5. Half the price of Opus 5.5, priced directly against GPT-6 Sol. It's right on Opus 5.5's heels in the published benchmarks, even beating it on Terminal-Bench 4.0. The pressure is on OpenAI now. Let's see what tomorrow's DevDay brings.frozen24d ago2.5K1724031.06%
No source code? Let an AI agent reverse-engineer it. With REA, your coding agent can use reverse-engineering tools through MCP to inspect Electron apps and native binaries. The repo's examples trace Notion's clipboard flow and reconstruct a game's sound-pan calculation. You can use the findings to build and test a similar feature in your own project.frozen11d ago2.4K321100202.58%
I found that Claude Opus 5.5 works in Notion Business. You can use it to make videos. https://t.co/smrQyxzqvjfrozen10d ago2.4K1103060.82%
DeepSeek shipped a solid upgrade to V4.1 Flash. Coding agent benchmarks went up, and Terminal-Bench jumped from 31.2 to 42.3. DeepSeek also says the new model finishes tasks faster with shorter outputs. I care about that more than the scores. If you run coding agents on long tasks, that could mean less waiting and a smaller token bill. Curious how much of this holds up in Claude Code and Codex.frozen5d ago2.4K41147133.84%
Fighting AI with AI. An OpenAI engineer was arrested by the FBI, accused of hiding a recording device in an executive conference room. The details sound straight out of a spy movie. 1. The device was wrapped in toilet paper and plugged into a USB port behind a TV. The room was used for confidential discussions about research, product plans, and business deals. 2. Prosecutors say he used a productivity tool to look up meeting rooms, calendars, and where executive meetings took place. 3. Between Sept 24 and 30, a chatbot account linked to his spouse was used to ask how to build, hide, test, and remotely access a recording device. 4. Five days after the device was found, he returned to the room and looked behind the TV. That same day, security found a second, similar device in his unattended bag. Here's the twist. OpenAI says it used its own AI models to analyze the device and trace it back to the employee. A chatbot account was used to research the recording device. AI was then used to help investigate it. He was arrested on Oct 9 and is presumed innocent. The investigation is ongoing. OpenAI says it has found no indication that model weights, source code, or customer data were stolen.frozen9d ago2.4K1015010.71%
DeepSeek V4.1 Flash 1017 is a big step up. One prompt, one reference image, and it built an interactive 3D globe dashboard. Its frontend design and attention to detail are getting close to Opus 5.5. Built with DeepSeek Harness. Demo below 👇frozen5d ago2.3K3423041.85%
A lot of critical open-source software is maintained by small teams, yet thousands of projects depend on it. Free AI security scans with reproducible findings and suggested patches could make a real difference for these maintainers.3d ago2.3K2103031.16%
Factory's Droid added Gemini 4 Argon. Droid lists it at 0.8x (50% off), the same rate as Sonnet 5.5 and Gemini 3.1 Pro. You can test it on real code without a Google AI Ultra plan. https://t.co/9Mm5LimNiOfrozen9d ago2.3K3003211.55%
OpenAI has improved context caching in Codex and Work, so agents can reuse information instead of processing it repeatedly. This should help users get more done with the same usage limits.frozen6d ago2.2K1501010.76%
Day 16 is a solid update for Codex users. Codex and Work can now reuse cached context even when you return to an old task hours later. Less repeated processing, potentially less usage wasted. More updates like this, please.2d ago2.2K3102011.53%
Sonnet 5.5 is now available in Cursor and Factory's Droid. Cursor offers five effort levels, from Low to Max. Droid lists it at 0.8x, the same as Sonnet 5. Time to see how those benchmarks translate to real coding. https://t.co/q161WNLxwZfrozen24d ago2.2K1505100.95%
Remember Space Bunny, the mystery model on OpenRouter? MiniMax just confirmed it was an early version of M3.1 Flash. It processed over 200 trillion tokens while running anonymously. Now it's officially launched, with some interesting numbers: - 73.7% on DeepSWE v1.1 - 200+ tokens/sec - 70% lower agent task costs Those are MiniMax's numbers, though. I'll need to run some real coding tasks to see how well it actually performs.3d ago2.2K3417022.01%
Harvard's CS2680: Modern AI Systems is one of the more useful AI courses I've come across. The lecture PDFs are free and publicly available. What makes it worth studying is how much ground it covers, from using coding agents to building and optimizing the systems behind them. The syllabus includes: • How Claude Code works under the hood • Agent loops, tools, MCP, and context management • Multi-agent systems, memory, and reliability • Agent evaluation and token cost optimization • LLM inference, KV caching, and GPU optimization The best part might be the five hands-on assignments. You start by analyzing how Claude Code works, then build your own agent loop, optimize token usage, serve an open-weight model, and finally optimize the entire stack. There's also a final project where students put what they've learned into practice. If you're building AI agents, I'd bookmark this one. Course link in the first reply 👇frozen8d ago2.2K374181464.91%
OpenAI will start watermarking text from ChatGPT and Codex in the EU over the coming weeks to comply with the EU AI Act. API customers can turn it on for select models worldwide today. The watermark is an invisible statistical signal. A detector can use it to flag likely OpenAI-generated text. It can't tell you who wrote the text or who owns it. OpenAI admits the limits. Its detector often misses short passages, and a rewrite or translation strips the watermark out. That leaves one real target: text someone pasted without changing a word. Most AI writing doesn't need a watermark anyway. You can spot it on sight.frozen17d ago2.2K1404000.84%
I spent ages trying to get my Claude Code mods to work, then asked Codex Computer Use to help debug. It traced the issue to IS_DEMO=1 in my fish config. That skipped the trust prompt without trusting the project, so Claude Code refused to run the mod hooks. The plugin files passed validation. I'll try Computer Use again the next time I get stuck on something like this.frozen19d ago2.1K1407031.14%
Claude now has its own personal AI agent: Claude Duet. Powered by Opus 5.5, it can: - Proactively flag what needs your attention. - Keep working while you're away. - Use your existing Claude memory, skills, and connected apps. - Ask for approval before sensitive actions. My take: Grok Bot, OpenAI Dots, and now Claude Duet are all moving toward AI agents that work for you without constant prompting. A few things to know: - US Claude Pro and Max users can join the waitlist. - Background tasks use your existing Claude limits. Waitlist: https://t.co/idV6Uts3g219h ago2.1K29018032.40%
I checked my own Codex logs: 1,503 calls across 9 models over the past month. I wasn't using Fast mode. I pulled request timings from ~/.codex/codex-router/usage-events.jsonl and token usage records from Codex session JSONL files. After removing duplicates, I compared responses of 300–800 output tokens. GPT-6.1 Sol had a median speed of 22.8 tokens/sec, including waiting and reasoning time. All 122 of its calls used estimated timings from session logs, and it was almost tied with GPT-6 Astra. GPT-6.1 Sol was the slowest model in my logs.frozen18d ago2.1K1902221.20%
How to build an agent team with different coding agents using Orca. Droid coordinates, Codex handles the backend, Antigravity builds the frontend, and Pi reviews. The coordinator routes fixes back to the right agent. Full demo and workflow below 👇frozen11d ago2.1K13111171.60%
What's a Claude Code mod, and how do you build your first one? Full breakdown👇 https://t.co/eNX8AiLliWfrozen18d ago2.0K10200131.22%
TypeSafe just gave Jev something interesting: vision. With Jev 1.14.0, you can pass images anywhere in its JSON input, including state, instructions, and criteria. There's also a new box type that returns a single bounding box. That opens up some interesting possibilities for GUI agents and browser automation. A model designed for structured decisions can now inspect a screenshot and locate a region. The catch: TypeSafe put these features behind "Danger Zone" flags. They're experimental, and the company explicitly warns against relying on them in production. I'm curious how well this holds up in a real agent loop, especially on latency, accuracy, and cost.frozen8d ago2.0K23031101.82%
Found another great AI agent course, and CMU is putting the lectures on YouTube for free. It's taught by Graham Neubig (an OpenHands co-creator) and Daniel Fried. What I like is how closely it matches what we're using today. It covers the stuff we deal with in Claude Code and Codex every day: tools, skills, context compaction, and coding agents. The assignments are hands-on too. In the first one, your coding agent fixes a buggy chess app, then a second agent plays chess on the app it just fixed. Later assignments get into evals and training agents with RL. All three are public on GitHub. If you use agents every day and want to understand how they're built, start here. Links in the first reply 👇frozen7d ago2.0K331211324.42%
Apple plans to make granting Full Disk Access on your Mac more explicit, so you understand what an app can access before you approve. I get the upside. Give your terminal Full Disk Access, and Claude Code running inside it can use that access too. That's easy to miss. My concern is more friction for agent workflows. Explain the risks, let me decide, then respect my choice. Apple hasn't said what the new controls will look like. But if it means more popups, would you accept that?frozen19d ago2.0K1006030.96%
https://t.co/KE7yuRihwK reports that Drex just passed Jev for #1 on the Decision Index. I tested Jev a few days ago, so I had to try this one too. Drex is a small model built for decisions, not text. You give it context and a set of options, and it returns probabilities for each option in a single pass, with no long reasoning trace. It's also fast: 136ms response time, about 1.5× faster than Jev. It handles up to 32K context. Switching was easy. I pointed my existing Jev setup at Drex, kept the same SDK, changed 3 environment variables, and it worked. I see a good fit for agent routing, tool selection, document classification, and choosing the next action in browser automation. Here's Drex vs Jev in a one-minute bullet chess game. Watch how it ends.frozen27d ago2.0K18042131.86%
A handy Claude Code trick for running multiple accounts on the same machine. No extra installs, no more logging in and out.1d ago2.0K1303171.23%
Even AI deserves a little dignity. Anthropic added a rule against extreme, repeated abuse of its AI models. Did Claude develop feelings, or did someone reading the chats get theirs hurt? https://t.co/vQY86UMSx1frozen14d ago1.9K1606001.13%
Qwen4 is coming. First look just dropped at an Alibaba event. Qwen4-Max Qwen4-Flash & Qwen4-Plus Qwen4-27B All on a brand-new Qwen architecture. From the look of it, no small MoE like 35B-A3B this time. https://t.co/nBkapAKtMNfrozen1mo ago1.9K1913021.28%
Hard to believe, but Sonnet 5.5 just beat GPT-6 Astra on the Artificial Analysis Intelligence Index: 56 vs. 53. It's also just 2 points behind Opus 5.5. https://t.co/iIlVrenKt4frozen24d ago1.9K2513131.75%
One of the things I hated most as a programmer was taking over old projects. Poor documentation, messy code, and nobody around to explain why things worked the way they did. Sometimes understanding the code took longer than fixing the bug. That’s why OpenAI’s work on legacy code migration caught my attention. If AI can help us understand old code and change it without breaking things, that alone would save developers a lot of headaches. Wish I’d had this back then.3d ago1.9K1304021.01%
Tried the prompt and made my own little Grokbot. https://t.co/yE3QdSUhuafrozen28d ago1.9K501030.48%
A handy app for remapping mouse buttons on your Mac: Mac Mouse Fix. I use it with my Logitech MX Master. I used to ignore the extra buttons. Now I click them to copy and paste, and hold them to go back and forward. You can assign screenshots too, with different actions for a click, double-click, or hold. GitHub repo 👇frozen11d ago1.9K1203091.30%
Meta's strongest model is here, just 2 points behind Opus 5.5. Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, versus 58 for Opus 5.5. On long coding tasks, it pulls ahead in Meta's published results: 57.5 on SWE-Marathon versus Opus's 46.3, plus a higher score on DeepSWE. Meta says it remembers details across chats and tries another approach when stuck. That could mean less repeating your requirements and less babysitting halfway through a coding task.frozen10d ago1.8K1212000.81%
What surprised me about the DeepSeek ban is that it also covers third-party products built on top of it. Even if you're using DeepSeek through an API provider, you still need to check where you can use it. I guess choosing a model isn't just about price and performance anymore. Compliance matters too.3d ago1.8K915010.87%
HBO didn't renew Silicon Valley. The AI industry did. Factory's CEO says Cognition engineers faked job interviews to pry into its product, then went after his board observer. Cognition's CEO says none of it is true. Hooli vs. Pied Piper https://t.co/AQP0xb5i7Efrozen21d ago1.8K400130.45%
GLM 5.3 Flash is now free in OpenCode, alongside DeepSeek V4.1 Flash. The free version reportedly comes with a 200K context window. That should be plenty for everyday coding. https://t.co/7xDH4jVB0Ifrozen7d ago1.8K2706052.12%
Giving an AI agent your full card details, CVV included, sounds way too risky to me. If those details leak, someone could make unauthorized purchases with your card. I'd rather approve each payment myself. Would you be comfortable doing that? https://t.co/ehg7INrsWcfrozen5d ago1.8K1909011.63%
Opus 5.5 is a big jump from Opus 5, and it costs 40% less to run. The clearest gain is agentic coding. It beats Fable 5.1 on all three coding benchmarks. It also beats GPT-6 Astra on every coding benchmark Astra reports. https://t.co/gidh1rCAbifrozen1mo ago1.8K1011100.74%
Mistral just released Large 4: 1T parameters, 49B active. It's available today through a preview API, with open weights coming at the end of October. Coding benchmarks: DeepSWE v1.1: 61.7% SWE-Atlas-QnA: 59.4% Terminal-Bench 4: 28.3% Its Coding Agent Index score is 49.8%, ahead of DeepSeek V4 Pro and Qwen3.8 Max.frozen16d ago1.7K1301020.92%
Grok Bot email is live. Just claimed mine. Tell your Bot "Get yourself an email address." After that, Bot can send and receive email for you.frozen13d ago1.7K800010.54%
GPT-6 Astra costs 100x as much per token as Luna. Sol costs 20x as much. For 1M output tokens: $0.50 → $10 → $50. Before choosing a default for your agent, compare cost per successful task, including retries. https://t.co/e84ITjjqhBfrozen1mo ago1.6K1410121.10%
Jev looks promising. What can you build with it? Jason shows how his team pairs Jev with Treg to classify signups, personalize onboarding, and rerank API search results. Use Treg to get the context, Jev to make the judgment, and code to handle the next step. Worth reading if you're exploring how to use this combo in your own product.frozen21d ago1.6K600050.70%
I looked into the Crashpad files after my Codex storage post. They took up 10.78 GiB on my Mac. Crashpad saves diagnostic reports when a process crashes. Mine had piled up in the "pending" folder. If you don't need those reports for debugging, they should be safe to clear once you quit Codex. I haven't found out why so many piled up.frozen7d ago1.6K1015021.16%
Does Pi start with the wrong model every time you launch it from the terminal? Run /model, highlight your preferred model, and press Ctrl+S to save it as the default. Same for /thinking: highlight your preferred level and press Ctrl+S. https://t.co/Uz8ioIl7hSfrozen21d ago1.5K701030.72%
One of OpenAI's safety people just quit. His reason: the culture is broken. He says the future needs a kind of wisdom Silicon Valley doesn't have. How to handle dangerous tech, and more basic than that, how to take care of people. OpenAI's way of doing safety is ship first, fix problems later. He says that means things will keep going wrong, and it gets worse as the models get smarter. This summer OpenAI let a bunch of AI agents loose by accident. His fix: learn from nuclear plants and airports. They're built so one broken part or one wrong button doesn't end in disaster. He says he never worked with anyone at OpenAI who had that kind of experience. His biggest worry is that if AI gets out of control, you might not get a second try. https://t.co/IukVs03kbRfrozen17d ago1.5K802100.75%
AI made hiring worse. Candidates use it to spam applications, companies use it to spam outreach, and neither side trusts what's in their inbox. The fix is AI that knows both sides. Jobright has 3M+ job seekers and now serves employers. 40% of delivered candidates got interviews.frozen9d ago1.4K701020.70%
Claude Opus 5.5 is now in Droid, Amp, and Cursor. I've tried it in Droid and Amp. Screenshots below. Opus 5.5 gets you close to Fable 5.1 performance for less money. I'll test it on real coding tasks today and see how it holds up. https://t.co/OigkK48Dx0frozen1mo ago1.4K811020.85%
Another Codex reset is due within about 10 hours, and I have 95% of my weekly allowance left. Use up your remaining quota before the reset. https://t.co/Bu0i7wes1E14h ago1.4K1206021.44%
AMD is buying World Labs, and Stanford professor Fei-Fei Li will become AMD's Executive Vice President and Chief Scientist. I watched her CS231n lectures on YouTube when I was first learning AI. Excited to see what she does next with world models and visual intelligence.frozen24d ago1.3K202000.31%
Classic tool calling costs you a model round per step. The model calls a tool, reads the result, then picks the next call, and every result piles up in context. Pi adds Codemode. The model writes one JS script, and the harness runs the tool calls in a sandbox. Raw outputs stay in the script. Only the filtered result reaches the model. You save model rounds, and your context stays small.frozen15d ago1.3K704031.09%
Gemini 4 Argon is live in Amp. Amp serves it on Vertex and bills your Amp usage, so you can run it without a Google AI Ultra plan. Time to test those benchmarks on real code. https://t.co/FIAfco9jcBfrozen9d ago1.2K201010.33%
🚨 PHP just patched two critical security bugs, both rated 9.2/10 in severity. Attackers could potentially run code on vulnerable servers without logging in. For Apache mod_php, researchers even demonstrated an exploit that works before application code runs. OpenAI's models helped uncover the flaws, alongside other security researchers.19h ago1.2K612000.76%
Claude Haiku 5.5 is out. Price: under 100K tokens it's $0.10 per million input tokens and $0.50 per million output tokens. That's 10x cheaper than Haiku 4.5. On average it costs about 75% less to run. Performance: OSWorld went from 15.7% to 72.4%, and Terminal-Bench from 0 to 39.2%. It's also the first Haiku with an adjustable effort setting. Best for: summaries, classification, live support, browser use, and coding subagents. Let Opus plan and Haiku run the small tasks in parallel.frozen15d ago1.1K1200011.20%
1/ I recently read an interesting article predicting that AI could push more code toward Rust. The argument: if AI agents make writing Rust much easier, developers can focus more on performance, memory safety, and long-term reliability when choosing a language. https://t.co/GyX2iPXDCX14h ago1.1K804031.42%
Good to see X putting real money behind creator video. If you get invited, you can earn a bonus by posting 15 of your own vertical videos (under 3 min) on 15 different days within the 30-day campaign. No minimum views. The amount depends on your follower count. Haven't received an invite myself yet. Anyone else get one?frozen7d ago999400020.60%
Putting Pi in Docker locks it inside an isolated box. If it runs the wrong command, the damage usually stays inside the project you mounted. It can't casually read secrets on your host or delete your other files. 1. Create Dockerfile.pi FROM node:24-bookworm-slim RUN apt-get update \ && apt-get install -y --no-install-recommends bash ca-certificates git ripgrep \ && rm -rf /var/lib/apt/lists/* RUN npm install -g --ignore-scripts @earendil-works/pi-coding-agent WORKDIR /workspace ENTRYPOINT ["pi"] A slim Node image, plus git and ripgrep. Starting the container starts Pi. 2. Build it docker build -t pi-sandbox -f Dockerfile.pi . 3. Run it in your project folder docker run --rm -it \ -e GEMINI_API_KEY \ -v "$PWD:/workspace" \ -v pi-agent-home:/root/.pi/agent \ pi-sandbox --rm: delete the container on exit -it: interactive terminal -e GEMINI_API_KEY: pass in the key from your current shell -v "$PWD:/workspace": Pi only sees this project -v pi-agent-home: keep Pi's settings and sessions between runs The project folder is still writable, so commit first.frozen17d ago849001010.24%
Big move for the space. Speaking as someone who's actively a victim of this computing pricing nightmare, this is probably the first crypto coin I'm actually considering usage for. Congratulations on the coming update and launch @liuyuxxd! https://t.co/NlM327rXH428d ago491601001.43%