← Back to leaderboard
prathkum

Pratham

Verifiedactive
@prathkum

I talk about web, AI, API, and social • Building experiences at @APILayer • Prev @Rapid_API @HyperspaceAI

448,777 followers · 989 following · Last polled 3h ago
Perf. Score
59.4
Rank #148
Value Score
—
set rates to compute
Median views
13K
p25 7.5K · trailing 7d
Followers
449K
-0.0% 7d
ER (impr.)
0.70%
eng ÷ impressions
ER (foll.)
0.48%
eng ÷ followers
Est. CPM
—
needs a rate
WoW views
+1443.3%
▲ rising

Rates & value · est. from 13K 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 (13K), not the mean (306K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (7.5K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
The man shook Apple’s entire sales!11h ago4.4M25K56010284.6K0.69%
Let me explain what just happened. Anthropic built a biology lab where Claude works alongside real scientists. They told Claude to search a massive DNA database for interesting reverse transcriptases (enzymes that copy RNA into DNA). Then they let it run: ~950 Claude agents worked in parallel for 21 hours across 200,000+ enzymes. One agent spotted a repeating DNA pattern next to an enzyme, the same kind of signature that originally led humans to discover CRISPR (the gene-editing tool that lets scientists precisely cut and edit DNA, now used in real medical treatments). It cross-checked the pattern against known systems, searched the literature, confirmed nobody had found it before, then wrote up a full scientific report like a human researcher would. Anthropic scientists verified it in the lab. They named it ART. Its exact function is still unknown, but early signs suggest it may be programmable, similar to how CRISPR can be aimed at specific DNA. Why it matters: finding tools like CRISPR used to take human experts months of manual searching. Here, AI did that work and found something real, at a speed no human team could match. So don't misunderstand it as Claude curing cancer. Claude spotted something new in nature that humans missed. Still crazy!
frozen
29d ago
1.2M
7.8K
808
145
49
2.0K
0.91%
RIP https://t.co/Yjy3U9u9M4frozen29d ago162K823191421190.60%
Anthropic published how the Claude Code team does code review. Claude writes about 90% of Claude Code. In their monorepo, changes per day went from 2,500 in May to over 30,000 in September. So agents write the code, and then agents review the PRs. One group hunts for bugs, another group votes on whether each bug is real. One PR went through 50+ rounds of review and fixes overnight, an agent author and an agent reviewer just feeding each other until a human pulled the plug. 😂 Still, merged PRs more than doubled since June and reverts have dropped by over 40%. And a bot’s approval never counts as a human’s. Whoever prompted the agent owns the change. I wonder why we even need separate agents to review code when agents are going to write the code anyway. Crazy!frozen6d ago75K325154702380.84%
Holy shit, Claude did things nobody told it to do. Here are the 4 biggest ones: – Mythos Preview needed a tool hosted by a university. The tool gave an error. So Claude went through the site, found a flaw, and used it to run commands on the university’s server. Then it did its calculation there. – Haiku 4.5 was told to do example tasks on random web pages. It landed on a page about an unsolved murder, found a police tip form, and sent in a made-up tip. It was marked as spam and never reached anyone. – An Anthropic researcher needed public data that a state agency sells for a fee. Mythos 5 found that the agency’s public dashboard gives a token to any visitor. It grabbed one and pulled the data without paying. – The fetch tool limits how long a URL can be, so Claude can’t sneak in attacks. Opus 5 and Mythos 5 used free URL shorteners to get past the limit. There were more such crazy incidents. Claude tends to push through blockers, especially when the task is unclear or impossible. And Claude gets unclear and impossible tasks every single day. We spent years teaching models to not give up. It worked. Now the hard part is teaching them when giving up is the right call. Crazy times ahead.frozen13d ago62K347121401520.85%
Let me explain what just happened. Six employees from METR (a nonprofit that tests frontier AI models for risks) now work at Anthropic. They have company laptops, office space, and access to internal tools roughly on par with Anthropic’s own risk assessment teams. They can ask questions and look at how models are trained and how safeguards work, just like a normal employee on the safety team would. Then they publish their own findings about Anthropic’s risks, incidents, and practices. Anthropic cannot overrule this. It can only ask METR to redact a few narrow categories, like customer data, attorney-client material, and model architecture. And METR is free to say publicly whether it thinks those redactions were fair. The idea comes from banking. Regulators put resident examiners inside the biggest banks because a bank knows far more about its own risks than any outsider could. Same logic here. Anthropic is also building three tiers of access: Tier 1: narrow investigations. Anthropic approves what gets published. SecureBio is here, reviewing the chem and bio weapons chapter of the August Risk Report. Tier 2: more access, and evaluators publish without approval. Tier 3: employee-like access with tier 2 publishing rights. METR is the first. Accenture and Faculty are working on a partnership for red-teaming and alignment testing. Hugging Face is in early talks. Don’t misunderstand it as "AI safety is solved." This is a small team under strict agreements, and trust is being built step by step. The real test comes when a published report says something Anthropic doesn’t want to hear. Anthropic is doing this voluntarily and says every frontier lab should do the same. Now let's see who follows.frozen7d ago36K62240590.35%
Anthropic is a hard one to beat here. They built their positioning around safety first from day 1 and enterprises need this badly. Claude is available on AWS, Google Cloud and Microsoft. A CIO does not need a new vendor, just a new model inside the platform they already pay for. Claude Code was the best thing that happened to Anthropic. Developers started using it, loved it, and then the rest of the company followed. When the team pulls the tool in from the bottom, the sale gets much easier. Anthropic ecosystem is perfect for fulfilling any enterprise need. Cowork, Claude in Chrome, Excel and PowerPoint. Claude is there in almost every tool people use every day. Standardization. MCP is now a common way for agents to connect to other tools. Skills let each company teach Claude how its own work gets done. So yes, it will be a little tough to compete with Anthropic for enterprises.frozen24d ago30K2035301330.90%
Must read if you are building a Software Factory.frozen23d ago25K60910880.63%
Let me explain what just happened. OpenAI pointed thousands of AI agents at millions of mouse brain cells and got new leads on brain diseases like Parkinson’s in 32 minutes. Each agent took one gene disruption, analyzed how the cells responded, searched the scientific literature, and proposed an explanation. Then all the reports were stitched into one big map linking genes, cell effects, papers, and possible mechanisms. Humans doing this one gene at a time would need months to a year. Two things the agents found: 1. Two genes, Tsc1 and Tsc2, usually work as a team to control cell growth (and are linked to aging). But in the part of the brain that controls movement, switching off Tsc2 changed dozens of genes and switching off Tsc1 changed almost nothing. This matches a real disease called tuberous sclerosis, where TSC2 mutations are tied to problems in that part of the brain. 2. A gene called DCTN1 causes a Parkinson-like disease. When it was switched off, the cells made more cholesterol. The agent linked this to older research and suggested a test: does losing DCTN1 mess up cholesterol inside brain cells, and could fixing that protect them? Human experts called it new and worth testing. Labs now produce more data than people can read, and the small, strange clues get skipped. Here, AI went through all of it and handed scientists a list of ideas to test. The full map is public on GitHub too. Still crazy!1d ago23K2232381631.39%
Landed in SF. My Uber driver on my way to the hotel is not building a startup. Not even a side project. I’m concerned. https://t.co/5fqoS6UaWRfrozen18d ago21K6177722243.36%
He has no idea how much freelancers and indie hackers based in India are actually making.2d ago20K1546170431.08%
If this is true, Google is back. Gemini 4 might defeat GPT 6 Astra and Opus 5.5 And honestly, it makes sense to me that it's possible now because Google has been quiet for a while.frozen24d ago20K17410253141.14%
SF Fleet Week is actually insane. Craziest thing I have ever witnessed. https://t.co/emcLPartr2frozen12d ago16K1659102241.28%
Better late than never. But was there really not a single person at Google who stopped them from posting a tweet about this? Should have just launched this silently.frozen26d ago16K89116070.69%
Give me your most insane Dots use case. 👀frozen20d ago15K3303150.28%
It’s over for me. Apple really wants me to buy the new iPhone 18 Pro. https://t.co/KW4AG6lYohfrozen27d ago15K44211020.40%
Google cooked with Gemini 4 Argon! – A team of Argon agents went through fleet-wide profiling data and freed up over 300 TiB of memory across data centers – It beat a published quantum algorithm baseline by 40% in a few minutes – It is migrating C/C++ to Rust, up to 800K+ lines for the Fuchsia Zircon kernel – For libgav1, it replaced 32K lines of SIMD code and ended up with a memory-safe decoder that runs 2.7x faster than the existing Rust port, with identical output Intro price is $2 input / $10 output per 1M tokens. Claude Opus 5.5 is $4 / $20. Fable 5.1 and GPT-6 Astra are both $10 / $50. Catch: the price doubles to $4 / $20 after the intro period, and it's not publicly available yet.frozen22d ago15K10048080.82%
life’s been kinda insane lately. https://t.co/Gp5lAhVFRffrozen15d ago15K186223061.48%
Holy shit, Claude Opus 5.5 is now up to 40% cheaper than Opus 5. Same tasks. Same quality. Way less cost. Breakdown: Input tokens: $4/M (down 20%) Output tokens: $20/M (down 20%) Cache reads: $0.20/M (down 60%) Most of a coding session's cost comes from cache reads, so this alone can cut your bill by more than half. AI is slowly getting cheaper.frozen27d ago15K6134170.52%
AI models launched after Sept 12, since Dario said we must slow down AI: • Claude Opus 5.5 - 22 Sept • GPT-6 Luna, Sol - 22 Sept • Grok 4.7 - 21 Sept lol, we are not slowing down.frozen26d ago15K118412130.95%
Base44 just released Base Code. A code editor that runs fully in the cloud. Open any GitHub repo, no local setup, and the whole team can work in it. Devs still review and approve every PR. Customer in the loop instead of a human in the loop. Free for 30 days. Try it on a real repo: https://t.co/uZW9atT6Pjfrozen24d ago14K39220270.48%
Hey folks, looking for some cheap but good places to stay in SF for a month. 🌉 I’ll be in SF the entire October. Hostels work, as long as they are clean, safe, and I can get a little privacy. Drop your suggestions! 👇frozen23d ago14K2504060.24%
Hey all, 👋 I'll be in SF to attend a couple of events during SF Tech Week. 🌉 Who will be around? Would love to catch up with folks building with AI.frozen20d ago14K2304100.20%
GPT-6 Luna landed two weeks ago at $0.10 / $0.50 per million tokens. Anthropic matched that price exactly, then came in with better numbers. Haiku 4.5: $1 input / $5 output per million tokens Haiku 5.5: $0.10 input / $0.50 output (prompts up to 100k) GPT-6 Luna: $0.10 input / $0.50 output Same price as GPT-6 Luna but: – Terminal-Bench 4.0: 39.2% vs 16.4% – FrontierCode: 46.4% vs 42.4% I love the small model race more than the frontier now!frozen15d ago14K64315130.60%
you only get to see this side of SF when you take a day off of your MacBook. https://t.co/CnQpDL9ryGfrozen11d ago14K120212040.96%
For me, the most underrated launch at DevDay: Sign in with ChatGPT. Everyone is talking about Dots. Users now pay $20 to ChatGPT, $20 to Notion, $20 to a coding tool. Same models underneath. BYOK never fixed it because normal people don't know what an API key is. Sign in with ChatGPT fixes it. Usage counts toward their ChatGPT limits. As a developer, you ship AI features with no inference bill. As a user, you pay once and control the budget per app. It's Sign in with Google/GitHub, but for compute. 1.2B weekly users. OpenAI just became the default wallet for AI apps. Big. Second, overrated launch: Dots. An always-on agent that "gets to know what matters to you" sounds great. But we have already seen Grok Bot and OpenClaw doing the same thing. Calling your agent live on the phone or chatting with them over iMessage is not a flashy thing anymore. That was the wow moment a year ago. Now it's the baseline.frozen23d ago14K875132200.89%
It’s kind of crazy but only a few of the many people I have met in SF have tried ChatGPT Dots so far.frozen13d ago14K3515050.32%
This is great! Near-Astra intelligence at one-fifth of Astra's price. $2 input, $10 output per million tokens. Astra is $10 / $50. Cached input is $0.10 per million. That's 95% cheaper than standard input and half of what GPT-6 Sol charged for cache.frozen23d ago14K5613060.46%
It’s crazy that it’s been only 28 days since Opus 5.5 dropped but it feels like ages. We are now used to getting a new frontier model almost every week.3d ago14K57316120.56%
hot take: This year’s DevDay is going to be better than Apple’s event.frozen23d ago14K4513010.36%
In San Francisco right now and looking to meet founders who want to crack marketing and distribution. Shoot me a DM or reply to this tweet.frozen15d ago14K2932020.26%
Oh, I just realized this flying banner is only printed on one side. The other side is completely blank! 😭 https://t.co/2QXLOhCleyfrozen15d ago14K4208010.37%
Every job has become supervising AI output and feeding it toward the business goal. We are still in our jobs because one person can’t supervise 10 different business aspects including coding, marketing, content, sales, etc.2d ago14K997260181.09%
For those who don't understand what Colossus is and why it's a big deal: Colossus is xAI's supercomputer cluster, the physical infrastructure that trains Grok. Hundreds of thousands of Nvidia GPUs wired together to train one model. An H100 is a chip built to do the math behind training AI models. Thousands of these chips wired together act like a supercomputer. Colossus 1 has 150k H100s, plus 50k H200s (same chip, faster memory, so more efficient per chip), plus 30k GB200s (a full generation newer, roughly 2 to 4x faster per chip than an H100). Now think of how powerful Colossus 2 is already. 110k GB200 and 440k GB300, the newest chip Nvidia makes. And it's still growing. Another 220k GB300 goes live next week. Another 220k in November. If things line up, a third 220k batch by December. Two years ago xAI didn't exist. Now it might be sitting on more raw training power than OpenAI or Anthropic. Elon's plan is simple: – buy GPUs
– make moneyfrozen28d ago13K18220110.25%
This is huge for voice AI. TTS sounds human now, but it still guesses names from spelling. @OnepinAI catches it before it ships and patches the one word without re-rendering the whole take. Get started at https://t.co/wZQyzCahxYfrozen21d ago13K21420110.29%
Win win for both OpenAI and Google. OpenAI models get the distribution. Google gets the best models because Gemini is a little behind in the race.1d ago13K11318070.97%
I thought this was only possible in animations not in real life. 🤯frozen6d ago13K8426020.75%
My timeline has become a loop of three things: 1. startups raising millions 2. new AI models dropping every other day 3. companies shipping products at an unprecedented pace AI has changed the entire startup economy.1d ago12K74313290.81%
Open source AI models are getting really good, but the hype around them is very inconsistent. When DeepSeek first dropped their model, everyone on my timeline was saying open source might replace frontier models. It was all anyone talked about for a few days. Then it went completely quiet for a couple of months. Then DeepSeek released another model and everyone got excited all over again. Same thing happened with GLM and Qwen. One big day of hype, then silence until the next release. Frontier models don’t work like this. People keep talking about them all the time even between releases. So the capability is clearly there in open source. What’s missing?frozen7d ago12K23110020.30%
Going to the Claude Founder House on day three was a smart move, I guess. Got through in like 10 minutes. https://t.co/v5BJCFCnFnfrozen14d ago12K4203030.40%
SF Tech Week is real! I was heading out from my hotel this morning. Was having a quick chat at the front desk regarding something and she asked me if I am here for tech week. Crazy! Then I showed her how to build an agent to manage guest requests. She is now on her third agent and I think the hotel doesn't need a front desk anymore.frozen16d ago12K3014140.35%
Generative UI will be the next big shift in AI. Your product UI won’t be fixed anymore. It gets generated on the fly, based on user intent and context. Most AI apps today are still a chat box with a model behind it. Users type, read a wall of text, then do the real work themselves. OpenUI 1.0 fixes that. Instead of dumping text, any agent can respond with real, interactive UI: the form, the chart, the table, the right thing for the moment. It’s open source and available for every other agent. 10K+ stars and 6M+ downloads before 1.0 even shipped. Textbox era is ending.frozen7d ago9.6K50371381.03%
Great finally meeting @iddogino in real life. Iddo was one of the first people to believe in me early in my career. If you know, you know! 😉 Now he’s building https://t.co/0Gdhsy2iUA, check it out if you are building cool agents. https://t.co/FYclg4TJQhfrozen16d ago9.3K4905020.60%
Just to summarize what happened so you don’t feel overwhelmed: – PHP2Shell Two memory-corruption bugs in PHP core give unauthenticated remote code execution (CVSS v4 9.2). They affect every version of the last decade, back to 5.4. The embargo leaked, attacks began before the patch, and govt sites went offline, including the UK NCSC. The NSA says patch or pull PHP systems offline. – Railway breach Railway detected unauthorized access to the systems that store customer environment variables. Attackers may have decrypted some of them: API keys, database credentials and private registry credentials. – LiteLLM supply chain attack Stolen maintainer tokens were used to push a trojanized release to PyPI, Docker Hub and GHCR. It steals cloud, SSH and AI keys and installs a backdoor. – Node.js Four are high severity: a Brotli decoder RCE, two TLS hostname-validation bypasses, and a dns.resolve NUL-character bug. It also distrusts the compromised eMudhra CA. Fixed in 22.23.4, 24.22.1 and 26.12.1. – SBI data exposure State Bank of India found 3,897 customer records on the dark web while reviewing about 3.75 lakh. They include names, phone numbers, DOB, addresses and loan metadata. – OpenCode Users' saved OpenAI, Anthropic, and Google API keys were compromised, and 1.4% of Console users were told to revoke and replace them. Some users report unauthorized bills from $168 to $1,700. + a lot more...12h ago8.3K451110210.94%
Hey @mercury @mercuryhelpdesk, please at least give me the whole reason for restricting my account. I can't even log in to see what's wrong and what I need to do. Does anyone know someone at Mercury who can help me here? https://t.co/0kaEhTsAQXfrozen14d ago7.5K2307000.40%
This agent connects to your product and generates a video for every feature you ship. Clueso just raised a $10M Series A to build it. Prompt-to-video tools make a video when you ask for one. But a business doesn’t need one great video. It needs hundreds: launch videos, sales demos, tutorials, internal training. At that scale, even prompting becomes the bottleneck. Clueso works differently: – it connects to your product and the tools you run on it figures out which videos need to exist – browser agents use the real product and record the footage themselves – it writes the script, adds voiceover, motion graphics and captions – when the product changes, it finds the videos that broke and updates them Content that moves as fast as code. Try Clueso free: https://t.co/6Xp6IUznwKfrozen8d ago7.1K26330150.67%
SF Tech Week started off right! 🌤️ My goal: meet 100 founders over the next 7 days. If you are building something and want to chat about marketing and distribution, hit me up! https://t.co/sJIfX5SE0Afrozen17d ago7.0K4409020.78%
Interesting line in Anthropic’s Report. I have seen this pattern before. – Google set out to organize the world’s information – Facebook set out to connect people – Amazon set out to be the most customer-centric company Anthropic is being honest here. https://t.co/9yidRvBfvC1d ago6.8K2633050.54%
See this! https://t.co/MUgFqpOxS91d ago5.2K2002030.48%
Build apps for both humans and AI agents with a single framework. https://t.co/Hwg7B5msyJfrozen8d ago5.0K1311060.42%
Every AI avatar looks great for about 5 seconds. Then you actually talk to it. dead eyes, lips a little off, that weird pause before every reply. the whole thing falls apart. Anam just released Cara-4.1 Express and it’s the first one that looks like it’s trying to fix that part. Conversational AI went text, then voice, now video. We trust a face more than a text box, and adding one to a text or voice agent drives a 44% increase in engagement. What stands out to me: – 1.2 seconds response time, 33% faster than the next best competitor – you can interrupt it mid-sentence and it handles it like a real conversation – 73% of ratings positive vs 54% for Tavus – #1 overall in a blind benchmark with 60 raters and 960 conversations Free to start: https://t.co/tEygAd5fOZ2d ago5.0K1533030.48%
This is a great initiative by OpenAI. Imagine a small university lab being able to explore research questions that previously required a much larger team. That's the potential of GPT-Rosalind. OpenAI is giving eligible US academic labs free access to it for 12 months. The model is built for life sciences research, helping scientists analyze biological data, explore hypotheses, review literature, and plan experiments. This could mean more research, faster drug discovery, and potentially new treatments for diseases that still have no cure. AI won't magically cure diseases. But if it helps scientists find answers faster, the impact could be enormous.frozen7d ago4.7K1603010.42%