Covering AI as it breaks. calling the hype when I see it
Joined August 2023
- Tweets41.5K
- Following316
- Followers1.5K
- Likes54.8K
Pinned Tweet
The last 24 hours in AI | Tue, Nov 3
• Anthropic's updated IPO filing says it has already used AI to build models whose jumps it missed at the time. It also warns it may soon hit its own bar for self-improving AI.
• Anthropic also changed Claude Code's instructions to refuse direct clones of existing video games. Original games and your own designs are still fine.
• OpenAI says its Georgia data center will give back up to 1,000 MW to the grid in heat waves or emergencies. That's enough for about 700,000 homes' AC.
• OpenAI is merging ChatGPT's Chat and Work tabs. ChatGPT and Codex will now draw from one token allowance, so some heavy users will hit limits sooner.
• The New York Times sued Brave, saying it scraped a 30 billion page index and sold it to AI labs. The suit names Anthropic, Perplexity, Mistral, xAI and Cohere as buyers.
• Elon Musk confirmed that Grok Bot weekly limits were cut on purpose. He said serving it cost too much relative to revenue. Some users jumped from 50% to 100% overnight.
• A research blog says GPT-6 Astra can read and edit model weights directly. It changed 0.016% of a small Qwen model and lifted its MATH-500 score from 7.9% to 48.1%.
• YC's Garry Tan says he spends about $100K to $200K a year on AI tokens. He runs agents for coding and work and shares 23 Claude Code skills on GitHub.
• 42AI came out of stealth with a $62.5M seed round. It says its model predicts how individual traders will react in volatile markets, and it is live at a few brokerages.
• Trajectory raised $45M to let lawyers, finance people and other experts train their own models without hiring a research team.
• Google launched its Dreambeans story app in India, its first market outside the US, with 13 local languages including Hindi, Tamil and Bengali.
My read: The same week a lab says AI may soon help build AI, two big products cut how much you can use. Self-improvement is cheap to claim in a filing. Paying to serve it is still the hard part.
teapaw.org/VT_zero1/status/211714…
The last 24 hours in AI | Mon, Nov 2
• OpenAI cut its new funding target to $20B, per the FT. The WSJ says it has $11B so far, and some big investors gave a tepid response.
• OpenAI is also planning a push after Tuesday's midterms, per Axios. Ideas include AI "speed limits" and calling data centers "AI compute campuses".
• Nvidia is in talks to guarantee billions in debt for Nscale's Anthropic data center in West Virginia. Anthropic signed $44.6B for the first building alone.
• xAI lost its bid to toss the NAACP lawsuit over its Memphis data center. A judge said the 35 gas turbines are not exempt as mobile sources.
• Apple's new CEO John Ternus told Wired the company won't ship AI for AI's sake. Separately, Apple Intelligence starts reaching China in the iOS 27.3 beta.
• OpenAI, Google, Microsoft and Anthropic agreed to honor one robots.txt line that opts a site out of AI training. Google says it starts in 2027.
• Mathematician Frank Calegari, on Terence Tao's blog, called OpenAI's 700 math papers a watershed. He also said many are badly written and some are wrong.
• Karpathy built a YouTube clone from HN picks in a day with Claude Opus 5.5. It sorted 40K comments down to 6,454 videos, and he wrote no code himself.
• Microsoft AI chief Mustafa Suleyman says AI must never become conscious. If signs show up, he wants development paused until they are removed.
• Vercel Labs open-sourced agent-desktop. It lets coding agents click and type in native apps on Mac, Windows and Linux by reading the accessibility tree.
• Intel fell about 28% on Friday after a weak Q4 outlook of $15B to $16B. AI data centers buying up memory are pushing PC prices higher.
My read: When the chip seller has to cosign the loans and the leader trims its raise, the buildout is running on nerve. Watch who ends up holding the debt.
teapaw.org/VT_zero1/status/211641…
VT retweeted
i thought i'd never post again
2.5 months ago
a doctor looked me in the eye and told me it was most likely cancer
found out i have a 14cm tumor and high CA markers
biopsy was not an option because the chances of spreading were high so i needed an open surgery
so i was wheeled into surgery not knowing who i'd be when i woke up
i lost my whole left ovary. then i waited weeks, scared out of my mind
it came back benign. NOT cancer
i cried so hard
still recovering after the surgery
so grateful that i get to live life, come back to work, say hi to my internet frens again
and yes i am alive, will be answering dms soon ❤️
VT retweeted
The last 24 hours in AI | Wed, Nov 4
• Anthropic released FlashProve, an open source prototype that proves which model wrote a reply without showing its weights. It's research only for now, capped at 1,024 tokens.
• Anthropic's new filing shows compute ate about $10B of its $17B Q3 revenue, or 58%. Half of that went to pricey short term capacity from smaller clouds.
• OpenAI says two staffers used GPT-6 Astra to design a working device in under a week, from case to circuit board to software. It still missed some factory limits.
• OpenAI more than quadrupled its bug bounty midpoints. A critical bug in ChatGPT, Codex or the API now pays up to $30K, and rare finds up to $172K with bonuses.
• Fal raised $1.3B at a $16B valuation, double its April price. The CEO says demand got so high this summer that they had to rate limit customers.
• Some data centers now run on retired jet engines while they wait years for grid power. ProEnergy's units give about 48 MW each, with more pollution than normal turbines.
• Marvell agreed to buy Lumentum in an all stock deal worth about $115B. The bet is that light, not copper, will move data inside AI data centers.
• Salesforce is buying Cresta, which builds voice agents for customer service, for a reported $2.7B. It plugs into Agentforce.
• Google told schools that turning off its AI tools means losing Google Search too. There's no admin switch to keep plain search and drop AI answers.
• A study found chatbots gave wrong or outdated voting info in 13.9% of tested answers, right before the midterms. Claude and Copilot made most of the clear errors.
• Apple will require an 18+ rating for dating apps and AI companions with romantic content. Developers must declare it starting Nov 16.
My read: A lab built a way to prove which model answered you in the same week its filing showed compute eating 58% of revenue. When serving costs this much, the urge to quietly swap in a cheaper model is real, so proof matters.
teapaw.org/VT_zero1/status/211750…
The last 24 hours in AI | Tue, Nov 3
• Anthropic's updated IPO filing says it has already used AI to build models whose jumps it missed at the time. It also warns it may soon hit its own bar for self-improving AI.
• Anthropic also changed Claude Code's instructions to refuse direct clones of existing video games. Original games and your own designs are still fine.
• OpenAI says its Georgia data center will give back up to 1,000 MW to the grid in heat waves or emergencies. That's enough for about 700,000 homes' AC.
• OpenAI is merging ChatGPT's Chat and Work tabs. ChatGPT and Codex will now draw from one token allowance, so some heavy users will hit limits sooner.
• The New York Times sued Brave, saying it scraped a 30 billion page index and sold it to AI labs. The suit names Anthropic, Perplexity, Mistral, xAI and Cohere as buyers.
• Elon Musk confirmed that Grok Bot weekly limits were cut on purpose. He said serving it cost too much relative to revenue. Some users jumped from 50% to 100% overnight.
• A research blog says GPT-6 Astra can read and edit model weights directly. It changed 0.016% of a small Qwen model and lifted its MATH-500 score from 7.9% to 48.1%.
• YC's Garry Tan says he spends about $100K to $200K a year on AI tokens. He runs agents for coding and work and shares 23 Claude Code skills on GitHub.
• 42AI came out of stealth with a $62.5M seed round. It says its model predicts how individual traders will react in volatile markets, and it is live at a few brokerages.
• Trajectory raised $45M to let lawyers, finance people and other experts train their own models without hiring a research team.
• Google launched its Dreambeans story app in India, its first market outside the US, with 13 local languages including Hindi, Tamil and Bengali.
My read: The same week a lab says AI may soon help build AI, two big products cut how much you can use. Self-improvement is cheap to claim in a filing. Paying to serve it is still the hard part.
teapaw.org/VT_zero1/status/211714…
The last 24 hours in AI | Wed, Nov 4
• Anthropic released FlashProve, an open source prototype that proves which model wrote a reply without showing its weights. It's research only for now, capped at 1,024 tokens.
• Anthropic's new filing shows compute ate about $10B of its $17B Q3 revenue, or 58%. Half of that went to pricey short term capacity from smaller clouds.
• OpenAI says two staffers used GPT-6 Astra to design a working device in under a week, from case to circuit board to software. It still missed some factory limits.
• OpenAI more than quadrupled its bug bounty midpoints. A critical bug in ChatGPT, Codex or the API now pays up to $30K, and rare finds up to $172K with bonuses.
• Fal raised $1.3B at a $16B valuation, double its April price. The CEO says demand got so high this summer that they had to rate limit customers.
• Some data centers now run on retired jet engines while they wait years for grid power. ProEnergy's units give about 48 MW each, with more pollution than normal turbines.
• Marvell agreed to buy Lumentum in an all stock deal worth about $115B. The bet is that light, not copper, will move data inside AI data centers.
• Salesforce is buying Cresta, which builds voice agents for customer service, for a reported $2.7B. It plugs into Agentforce.
• Google told schools that turning off its AI tools means losing Google Search too. There's no admin switch to keep plain search and drop AI answers.
• A study found chatbots gave wrong or outdated voting info in 13.9% of tested answers, right before the midterms. Claude and Copilot made most of the clear errors.
• Apple will require an 18+ rating for dating apps and AI companions with romantic content. Developers must declare it starting Nov 16.
My read: A lab built a way to prove which model answered you in the same week its filing showed compute eating 58% of revenue. When serving costs this much, the urge to quietly swap in a cheaper model is real, so proof matters.
teapaw.org/VT_zero1/status/211750…
The last 24 hours in AI | Tue, Nov 3
• Anthropic's updated IPO filing says it has already used AI to build models whose jumps it missed at the time. It also warns it may soon hit its own bar for self-improving AI.
• Anthropic also changed Claude Code's instructions to refuse direct clones of existing video games. Original games and your own designs are still fine.
• OpenAI says its Georgia data center will give back up to 1,000 MW to the grid in heat waves or emergencies. That's enough for about 700,000 homes' AC.
• OpenAI is merging ChatGPT's Chat and Work tabs. ChatGPT and Codex will now draw from one token allowance, so some heavy users will hit limits sooner.
• The New York Times sued Brave, saying it scraped a 30 billion page index and sold it to AI labs. The suit names Anthropic, Perplexity, Mistral, xAI and Cohere as buyers.
• Elon Musk confirmed that Grok Bot weekly limits were cut on purpose. He said serving it cost too much relative to revenue. Some users jumped from 50% to 100% overnight.
• A research blog says GPT-6 Astra can read and edit model weights directly. It changed 0.016% of a small Qwen model and lifted its MATH-500 score from 7.9% to 48.1%.
• YC's Garry Tan says he spends about $100K to $200K a year on AI tokens. He runs agents for coding and work and shares 23 Claude Code skills on GitHub.
• 42AI came out of stealth with a $62.5M seed round. It says its model predicts how individual traders will react in volatile markets, and it is live at a few brokerages.
• Trajectory raised $45M to let lawyers, finance people and other experts train their own models without hiring a research team.
• Google launched its Dreambeans story app in India, its first market outside the US, with 13 local languages including Hindi, Tamil and Bengali.
My read: The same week a lab says AI may soon help build AI, two big products cut how much you can use. Self-improvement is cheap to claim in a filing. Paying to serve it is still the hard part.
teapaw.org/VT_zero1/status/211714…
VT retweeted
Top AI repos on GitHub today | Wed, Nov 4
• vercel-labs/agent-desktop: A command line tool that lets an AI agent use your computer. It can click around and read what is on screen.
github.com/vercel-labs/agent…
• Donchitos/Claude-Code-Game-Studios: Turns Claude Code into a small game studio, with 49 agents that each play a role.
github.com/Donchitos/Claude-…
• JustVugg/colibri: Runs huge MoE models on normal hardware. Busy parts sit on the GPU, the rest wait in RAM or on disk.
github.com/JustVugg/colibri
• p-e-w/heretic: Changes how a model behaves without retraining it. Less refusing, less flattery.
github.com/p-e-w/heretic
• FlashML-org/FreeToken: Serves very large models from your own desktop, fast.
github.com/FlashML-org/FreeT…
• MakazhanAlpamys/Soup: Fine-tune a model from one config file. It trains an 8B model on a 4 GB laptop GPU.
github.com/MakazhanAlpamys/S…
My pick: colibri. Three of today's six are about running big models on hardware you already own. People are tired of renting every token.
teapaw.org/VT_zero1/status/211598…
Top AI repos on GitHub today | Fri, Oct 30
• pbakaus/impeccable: Gives your coding agent a design vocabulary. You tell it to polish or go bolder, and it reviews its own UI work.
github.com/pbakaus/impeccabl…
• vectorize-io/hindsight: Memory for agents that actually learns from past runs, instead of just storing old chats.
github.com/vectorize-io/hind…
• obra/superpowers: A skills framework that teaches coding agents a real way to build software, step by step.
github.com/obra/superpowers
• thedotmack/claude-mem: Saves what your agent did in one session so the next session starts with that context.
github.com/thedotmack/claude…
• baidu/Unlimited-OCR: Baidu's OCR model that reads long documents in one pass.
github.com/baidu/Unlimited-O…
• CopilotKit/OpenBot: Open-source AI coworkers. Each one gets its own computer with a browser, files and tools.
github.com/CopilotKit/OpenBo…
My pick: Hindsight. Two of today's six are about agent memory. That tells you where the pain is. Agents are smart enough now, they just forget everything by tomorrow.
Top AI repos on GitHub today | Wed, Nov 4
• vercel-labs/agent-desktop: A command line tool that lets an AI agent use your computer. It can click around and read what is on screen.
github.com/vercel-labs/agent…
• Donchitos/Claude-Code-Game-Studios: Turns Claude Code into a small game studio, with 49 agents that each play a role.
github.com/Donchitos/Claude-…
• JustVugg/colibri: Runs huge MoE models on normal hardware. Busy parts sit on the GPU, the rest wait in RAM or on disk.
github.com/JustVugg/colibri
• p-e-w/heretic: Changes how a model behaves without retraining it. Less refusing, less flattery.
github.com/p-e-w/heretic
• FlashML-org/FreeToken: Serves very large models from your own desktop, fast.
github.com/FlashML-org/FreeT…
• MakazhanAlpamys/Soup: Fine-tune a model from one config file. It trains an 8B model on a 4 GB laptop GPU.
github.com/MakazhanAlpamys/S…
My pick: colibri. Three of today's six are about running big models on hardware you already own. People are tired of renting every token.
teapaw.org/VT_zero1/status/211598…
Top AI repos on GitHub today | Fri, Oct 30
• pbakaus/impeccable: Gives your coding agent a design vocabulary. You tell it to polish or go bolder, and it reviews its own UI work.
github.com/pbakaus/impeccabl…
• vectorize-io/hindsight: Memory for agents that actually learns from past runs, instead of just storing old chats.
github.com/vectorize-io/hind…
• obra/superpowers: A skills framework that teaches coding agents a real way to build software, step by step.
github.com/obra/superpowers
• thedotmack/claude-mem: Saves what your agent did in one session so the next session starts with that context.
github.com/thedotmack/claude…
• baidu/Unlimited-OCR: Baidu's OCR model that reads long documents in one pass.
github.com/baidu/Unlimited-O…
• CopilotKit/OpenBot: Open-source AI coworkers. Each one gets its own computer with a browser, files and tools.
github.com/CopilotKit/OpenBo…
My pick: Hindsight. Two of today's six are about agent memory. That tells you where the pain is. Agents are smart enough now, they just forget everything by tomorrow.
VT retweeted
John Goodman admits he was at "a really low point" while making Bringing Out the Dead with Scorsese and Nic Cage.
He was drinking and couldn't find the character, riding around in ambulances with Marty from six at night to six in the morning. Then his wife got appendicitis while visiting, so he'd go straight from set to the hospital until he started breaking down. He says he just wasn't there for anybody. He was close to 400 pounds, and he thought vodka hid it, but the makeup people could tell from his red cheeks.
When he says "I was on my way out," you can tell he means it.
ft: @joshuahorowitz Happy Sad Confused on youtube
VT retweeted
Josh Brolin says he made more money in four years of day trading stocks than in 20 years of acting.
He got $100K for No Country for Old Men and says he cleared about $30K of it. While he was making it, he sent in an audition tape for American Gangster with his 16-year-old son reading Denzel's part. He says that was the last audition he ever did.
Everyone assumes a Coen brothers movie makes you rich. For him it barely made a decent year.
ft: @TheoVon This Past Weekend on youtube
VT retweeted
The last 24 hours in AI | Tue, Nov 3
• Anthropic's updated IPO filing says it has already used AI to build models whose jumps it missed at the time. It also warns it may soon hit its own bar for self-improving AI.
• Anthropic also changed Claude Code's instructions to refuse direct clones of existing video games. Original games and your own designs are still fine.
• OpenAI says its Georgia data center will give back up to 1,000 MW to the grid in heat waves or emergencies. That's enough for about 700,000 homes' AC.
• OpenAI is merging ChatGPT's Chat and Work tabs. ChatGPT and Codex will now draw from one token allowance, so some heavy users will hit limits sooner.
• The New York Times sued Brave, saying it scraped a 30 billion page index and sold it to AI labs. The suit names Anthropic, Perplexity, Mistral, xAI and Cohere as buyers.
• Elon Musk confirmed that Grok Bot weekly limits were cut on purpose. He said serving it cost too much relative to revenue. Some users jumped from 50% to 100% overnight.
• A research blog says GPT-6 Astra can read and edit model weights directly. It changed 0.016% of a small Qwen model and lifted its MATH-500 score from 7.9% to 48.1%.
• YC's Garry Tan says he spends about $100K to $200K a year on AI tokens. He runs agents for coding and work and shares 23 Claude Code skills on GitHub.
• 42AI came out of stealth with a $62.5M seed round. It says its model predicts how individual traders will react in volatile markets, and it is live at a few brokerages.
• Trajectory raised $45M to let lawyers, finance people and other experts train their own models without hiring a research team.
• Google launched its Dreambeans story app in India, its first market outside the US, with 13 local languages including Hindi, Tamil and Bengali.
My read: The same week a lab says AI may soon help build AI, two big products cut how much you can use. Self-improvement is cheap to claim in a filing. Paying to serve it is still the hard part.
teapaw.org/VT_zero1/status/211714…
The last 24 hours in AI | Mon, Nov 2
• OpenAI cut its new funding target to $20B, per the FT. The WSJ says it has $11B so far, and some big investors gave a tepid response.
• OpenAI is also planning a push after Tuesday's midterms, per Axios. Ideas include AI "speed limits" and calling data centers "AI compute campuses".
• Nvidia is in talks to guarantee billions in debt for Nscale's Anthropic data center in West Virginia. Anthropic signed $44.6B for the first building alone.
• xAI lost its bid to toss the NAACP lawsuit over its Memphis data center. A judge said the 35 gas turbines are not exempt as mobile sources.
• Apple's new CEO John Ternus told Wired the company won't ship AI for AI's sake. Separately, Apple Intelligence starts reaching China in the iOS 27.3 beta.
• OpenAI, Google, Microsoft and Anthropic agreed to honor one robots.txt line that opts a site out of AI training. Google says it starts in 2027.
• Mathematician Frank Calegari, on Terence Tao's blog, called OpenAI's 700 math papers a watershed. He also said many are badly written and some are wrong.
• Karpathy built a YouTube clone from HN picks in a day with Claude Opus 5.5. It sorted 40K comments down to 6,454 videos, and he wrote no code himself.
• Microsoft AI chief Mustafa Suleyman says AI must never become conscious. If signs show up, he wants development paused until they are removed.
• Vercel Labs open-sourced agent-desktop. It lets coding agents click and type in native apps on Mac, Windows and Linux by reading the accessibility tree.
• Intel fell about 28% on Friday after a weak Q4 outlook of $15B to $16B. AI data centers buying up memory are pushing PC prices higher.
My read: When the chip seller has to cosign the loans and the leader trims its raise, the buildout is running on nerve. Watch who ends up holding the debt.
teapaw.org/VT_zero1/status/211641…
VT retweeted
William Shatner says he and his youngest daughter were diagnosed with stage 4 cancer at about the same time.
He had melanoma, she had breast cancer. They both landed with doctors who knew exactly what they were dealing with, and he says ten years earlier the same news would've come with two or three months to live. It didn't bother him, because he's been lucky his whole life. His daughter was much younger, and she was terrified of dying.
He's 95 and he was the one who wasn't scared. Hard to hear that and not think about your own family.
ft: @esquire Esquire on youtube
VT retweeted
ROPES just passed 100 followers. Thank you for climbing with us.
Season 0 starts today on @Enter_Elysium testnet:
Points for every ranked seat
Weekly boards, reset every Monday
5 leagues, Base Camp to Summit
13 trophies and a daily streak
playropes.xyz/
VT retweeted
The last 24 hours in AI | Mon, Nov 2
• OpenAI cut its new funding target to $20B, per the FT. The WSJ says it has $11B so far, and some big investors gave a tepid response.
• OpenAI is also planning a push after Tuesday's midterms, per Axios. Ideas include AI "speed limits" and calling data centers "AI compute campuses".
• Nvidia is in talks to guarantee billions in debt for Nscale's Anthropic data center in West Virginia. Anthropic signed $44.6B for the first building alone.
• xAI lost its bid to toss the NAACP lawsuit over its Memphis data center. A judge said the 35 gas turbines are not exempt as mobile sources.
• Apple's new CEO John Ternus told Wired the company won't ship AI for AI's sake. Separately, Apple Intelligence starts reaching China in the iOS 27.3 beta.
• OpenAI, Google, Microsoft and Anthropic agreed to honor one robots.txt line that opts a site out of AI training. Google says it starts in 2027.
• Mathematician Frank Calegari, on Terence Tao's blog, called OpenAI's 700 math papers a watershed. He also said many are badly written and some are wrong.
• Karpathy built a YouTube clone from HN picks in a day with Claude Opus 5.5. It sorted 40K comments down to 6,454 videos, and he wrote no code himself.
• Microsoft AI chief Mustafa Suleyman says AI must never become conscious. If signs show up, he wants development paused until they are removed.
• Vercel Labs open-sourced agent-desktop. It lets coding agents click and type in native apps on Mac, Windows and Linux by reading the accessibility tree.
• Intel fell about 28% on Friday after a weak Q4 outlook of $15B to $16B. AI data centers buying up memory are pushing PC prices higher.
My read: When the chip seller has to cosign the loans and the leader trims its raise, the buildout is running on nerve. Watch who ends up holding the debt.
teapaw.org/VT_zero1/status/211641…
The last 24 hours in AI | Sat, Oct 31
• Anthropic renamed its top plans. Max 5x is now Max at $100, and Max 20x is now Ultra at $200 with no 5-hour limit for verified accounts.
• Anthropic rewrote its Messages API in Rust using Claude, with no Rust developers on the team. CPU use fell 40 to 50%.
• Anthropic also used thousands of Claude agents to red-team for subtle misalignment, work that people used to do by hand.
• OpenAI launched Private Intelligence, which gives businesses tighter data controls on its top models.
• Sam Altman says OpenAI is holding many more math discoveries than it has made public, per The Verge.
• OpenAI's Codex now lets you switch accounts on mobile, with desktop next. Its lead hinted at limit resets in the next two days.
• Google ran a prompt on a TPU inside a prototype satellite for Project Suncatcher. It sent back a haiku from orbit.
• Reddit's CEO says Google AI Mode sends it no traffic. Search referrals fell by over 1M a week in October.
• Mistral's Le Chonk, an update to Large 4, ranked second among open models on Requesty's coding benchmark.
• xAI shipped the first stable Grok Bot. Testers spotted a Tasks hub and texting support in the works.
• The NYT reports a Democratic House would likely drop the hands-off approach to AI companies.
My read: The Reddit number is the one to watch. If AI answers keep eating the clicks, the sites these models learned from stop getting paid to exist.
teapaw.org/VT_zero1/status/211605…
