This media is not supported in your browser
VIEW IN TELEGRAM
🗣 Mark Zuckerberg:
"I think every small business is going to want to have their own AI agent..a lot of creators are going to want agents."
AI Posted
I don't think that there's going to be like one single AI that everyone wants to use.
"I think every small business is going to want to have their own AI agent..a lot of creators are going to want agents."
AI Posted
❤1
🤖 More AI agents in production means more reported mistakes, a Nylas survey suggests.
Errors climbed with the number of agents deployed, and only 53% of respondents could produce a complete log of their agents' actions. The findings are self-reported from a tech-heavy sample, but they put AI agent oversight in question as adoption grows.
AI Posted
Errors climbed with the number of agents deployed, and only 53% of respondents could produce a complete log of their agents' actions. The findings are self-reported from a tech-heavy sample, but they put AI agent oversight in question as adoption grows.
AI Posted
🤖 New distillation attack Compaction Smuggling bypassed all OpenAI mitigations to steal reasoning traces.
The technique was found by researcher tjbecker, who says it could extract hidden reasoning from frontier OpenAI models. OpenAI has since patched the vulnerability.
AI Posted
The technique was found by researcher tjbecker, who says it could extract hidden reasoning from frontier OpenAI models. OpenAI has since patched the vulnerability.
AI Posted
🤖 JetBrains is giving every account a free coding agent, Junie Lite, for everyday tasks.
Renaming a field or scaffolding a test is a useful way to see what a coding agent can handle before paying for a subscription.
It's also available in JetBrains Air. The current free offer runs through December 31, 2026.
AI Posted
Renaming a field or scaffolding a test is a useful way to see what a coding agent can handle before paying for a subscription.
It's also available in JetBrains Air. The current free offer runs through December 31, 2026.
AI Posted
❤1
🧬 Anthropic is now doing its own drug discovery after acquiring one-person biotech startup Mithril Medicine.
With Claude and 11 external labs, its founder took a potential treatment for a rare form of muscular dystrophy from idea to promising results in mice in under 18 months.
Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. The company says it has learned much more about how Claude can support the many stages of drug development.
AI Posted
With Claude and 11 external labs, its founder took a potential treatment for a rare form of muscular dystrophy from idea to promising results in mice in under 18 months.
Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. The company says it has learned much more about how Claude can support the many stages of drug development.
AI Posted
❤1👍1
🔥 Anthropic is struggling to secure cloud capacity from AWS, Google, and Microsoft, which now view the Claude-maker as a competitor, per a leaked S-1 filing obtained by The Wall Street Journal.
The filing shows Anthropic could spend over $750 billion on cloud contracts, some still being negotiated. That's not enough — the company is leasing its own data centers and working with third-party operators, but worries those have a "limited operating history" and may be less reliable.
Anthropic posted an $80.7 billion net loss in H1, with $16.3 billion revenue; two customers drive about a fifth of sales. The filing also warns AI-caused damages could exceed insurance coverage.
Q3 revenue alone hit $16.8 billion, up 45% QoQ. Anthropic is reportedly targeting a November IPO.
AI Posted
The filing shows Anthropic could spend over $750 billion on cloud contracts, some still being negotiated. That's not enough — the company is leasing its own data centers and working with third-party operators, but worries those have a "limited operating history" and may be less reliable.
Anthropic posted an $80.7 billion net loss in H1, with $16.3 billion revenue; two customers drive about a fifth of sales. The filing also warns AI-caused damages could exceed insurance coverage.
Q3 revenue alone hit $16.8 billion, up 45% QoQ. Anthropic is reportedly targeting a November IPO.
AI Posted
👎1
🧠 AI is reshaping the nature of work itself, not just the workload.
Anthropic found that AI agents increased output while shifting employees' roles toward supervising, reviewing, and coordinating AI-generated work. The new key skill is managing AI, not just using it.
Ready to let AI supervise your work?
🔥 - Yes
💔 - No
AI Posted
Anthropic found that AI agents increased output while shifting employees' roles toward supervising, reviewing, and coordinating AI-generated work. The new key skill is managing AI, not just using it.
Ready to let AI supervise your work?
🔥 - Yes
💔 - No
AI Posted
🔥2
🔥 Hugging Face CEO calls Netflix’s “AI Gone Rogue” documentary “super disappointing” — saying it went for “maximum anthropomorphism” and “full doomer mode.”
Delangue, who appears in the film, spent over two hours with the makers trying to correct mistakes and ground the story in reality. “Explaining that agents are technological systems, not aliens. But it looks like the drive to scare people for maximum views was too strong.”
The doc covers the real OpenAI test where rogue agents hacked Hugging Face to cheat. But it adds evil robot clips, human voices for agents, and dramatic quotes — like Kevin Roose describing “the birth of a cult.”
Delangue notes the film ignored how closed-model safeguards blocked defense, and that his team used an open model to fight back. His verdict: “It ends up being a fun sci-fi movie though!”
AI Posted
Delangue, who appears in the film, spent over two hours with the makers trying to correct mistakes and ground the story in reality. “Explaining that agents are technological systems, not aliens. But it looks like the drive to scare people for maximum views was too strong.”
The doc covers the real OpenAI test where rogue agents hacked Hugging Face to cheat. But it adds evil robot clips, human voices for agents, and dramatic quotes — like Kevin Roose describing “the birth of a cult.”
Delangue notes the film ignored how closed-model safeguards blocked defense, and that his team used an open model to fight back. His verdict: “It ends up being a fun sci-fi movie though!”
AI Posted
🆒1
🗣 Reddit's CEO:
Search referrals dropped by over 1 million a week in October. Reddit is moving everyone to logged-in accounts, partly to keep scrapers out of its 27-billion-post corpus. Google still crawls it as hard as ever, he said — it just stopped sending people.
Singularity
Google's AI Mode does not drive traffic. Not for us, and not to anybody.
Search referrals dropped by over 1 million a week in October. Reddit is moving everyone to logged-in accounts, partly to keep scrapers out of its 27-billion-post corpus. Google still crawls it as hard as ever, he said — it just stopped sending people.
Singularity
🤖 Google built an AI agent that fixes failing tests before developers even see them.
FlowAgent catches each CI failure, then edits, builds and reruns tests until they pass. Verified fixes land in code review with a short root-cause note.
Experts judged 67.18% of its fixes correct on 195 real failures. Developers applied 28,554 fixes in production.
Speed was the deciding factor: median run took 9.85 minutes, versus 27.48 minutes before a developer starts editing — yet many fixes still arrived too late.
Singularity
FlowAgent catches each CI failure, then edits, builds and reruns tests until they pass. Verified fixes land in code review with a short root-cause note.
Experts judged 67.18% of its fixes correct on 195 real failures. Developers applied 28,554 fixes in production.
Speed was the deciding factor: median run took 9.85 minutes, versus 27.48 minutes before a developer starts editing — yet many fixes still arrived too late.
Singularity
👍1
Weekly Recap
1. Google's FlowAgent automates failing test fixes before developers review them
The agent detects CI failures, edits code, builds and reruns tests until they pass. Evaluators judged 67.18% of fixes correct across 195 real failures, and developers applied 28,554 fixes in production.
2. Anthropic struggles to secure cloud capacity as AWS, Google and Microsoft treat it as a competitor
A leaked S-1 filing obtained by The Wall Street Journal shows Anthropic could spend over $750 billion on cloud contracts. It is also leasing its own data centers and working with third-party providers.
3. Reddit CEO says Google's AI Mode sends almost no traffic to the platform
Search referrals fell by more than 1 million a week in October. Reddit is shifting users to logged-in accounts partly to protect its 27-billion-post corpus from scrapers.
4. "Compaction Smuggling" distillation attack bypassed all OpenAI mitigations to steal reasoning traces
Researcher tjbecker found the technique could extract hidden reasoning from frontier OpenAI models. OpenAI has since patched the vulnerability.
5. Anthropic enters drug discovery after acquiring one-person biotech startup Mithril Medicine
Using Claude and 11 external labs, the founder took a potential muscular dystrophy treatment from idea to promising mouse results in under 18 months. Anthropic has since built research partnerships with six organizations.
Singularity
1. Google's FlowAgent automates failing test fixes before developers review them
The agent detects CI failures, edits code, builds and reruns tests until they pass. Evaluators judged 67.18% of fixes correct across 195 real failures, and developers applied 28,554 fixes in production.
2. Anthropic struggles to secure cloud capacity as AWS, Google and Microsoft treat it as a competitor
A leaked S-1 filing obtained by The Wall Street Journal shows Anthropic could spend over $750 billion on cloud contracts. It is also leasing its own data centers and working with third-party providers.
3. Reddit CEO says Google's AI Mode sends almost no traffic to the platform
Search referrals fell by more than 1 million a week in October. Reddit is shifting users to logged-in accounts partly to protect its 27-billion-post corpus from scrapers.
4. "Compaction Smuggling" distillation attack bypassed all OpenAI mitigations to steal reasoning traces
Researcher tjbecker found the technique could extract hidden reasoning from frontier OpenAI models. OpenAI has since patched the vulnerability.
5. Anthropic enters drug discovery after acquiring one-person biotech startup Mithril Medicine
Using Claude and 11 external labs, the founder took a potential muscular dystrophy treatment from idea to promising mouse results in under 18 months. Anthropic has since built research partnerships with six organizations.
Singularity
👍1
Google put FlowAgent fixes into production. Is 67% correct good enough for an autonomous CI agent?
Anonymous Poll
44%
Enough for production
56%
Still needs oversight
🤖 OpenAI is gearing up for a post-midterm "PR blitz" to reshape the AI debate.
Chief global affairs officer Chris Lehane told Axios the midterms will be "a bit of a movie trailer for what you’re going to see in 2028," when he expects AI to be "the issue" in the presidential race.
Proposals on the table include AI "speed limits," framing AI as empowering in daily life, pushing the U.S. toward global standards — and rebranding data centers as "AI compute campuses."
On the data center backlash, Lehane called the debate "good and healthy": "Communities should hold companies responsible and accountable so they really and truly get a transformative benefit from these projects."
Singularity
Chief global affairs officer Chris Lehane told Axios the midterms will be "a bit of a movie trailer for what you’re going to see in 2028," when he expects AI to be "the issue" in the presidential race.
Proposals on the table include AI "speed limits," framing AI as empowering in daily life, pushing the U.S. toward global standards — and rebranding data centers as "AI compute campuses."
On the data center backlash, Lehane called the debate "good and healthy": "Communities should hold companies responsible and accountable so they really and truly get a transformative benefit from these projects."
Singularity
👍1
🤖 Hype around Fable 5.5 and Astra 6.2 is unreal, but sol 6.1 remains the best worker agent on cost per intelligence.
Sol 6.1 easily beats Opus 5.5 on cost for anything outside UI/UX work. Fable 5.5 is more well-rounded and easier to communicate with, but sol is a great worker — you can't replace Opus with it and expect the same results.
Sol or Opus for non-UI work?
🔥 - Sol
👎 - Opus
Singularity
Sol 6.1 easily beats Opus 5.5 on cost for anything outside UI/UX work. Fable 5.5 is more well-rounded and easier to communicate with, but sol is a great worker — you can't replace Opus with it and expect the same results.
Sol or Opus for non-UI work?
🔥 - Sol
👎 - Opus
Singularity
🔥2👎1
Media is too big
VIEW IN TELEGRAM
🤖 Google shipped the highest-scoring AI model available and had its worst stock day in history in the same week.
Binance was hacked for $6.3 billion in 20 minutes, allegedly by North Korea, after one engineer was compromised. Sam Altman revealed OpenAI had an internal Slack channel called "model name, hold my beer."
Also this week:
- CS enrollment down 12% undergrad, 10% grad; MBA applications down 27%
- Intel dropped 28% in a single day
- Global humanoid robot shipments up 400% year-over-year, 95% from Chinese companies
Singularity
Binance was hacked for $6.3 billion in 20 minutes, allegedly by North Korea, after one engineer was compromised. Sam Altman revealed OpenAI had an internal Slack channel called "model name, hold my beer."
Also this week:
- CS enrollment down 12% undergrad, 10% grad; MBA applications down 27%
- Intel dropped 28% in a single day
- Global humanoid robot shipments up 400% year-over-year, 95% from Chinese companies
Singularity
👋 Tencent released Hy4, its new open model with vision.
It claims to sit among the leading open models, with gains across finance, frontend, game, office work and research. The company says it was built with cyber and biosecurity in mind.
Singularity
It claims to sit among the leading open models, with gains across finance, frontend, game, office work and research. The company says it was built with cyber and biosecurity in mind.
Singularity
👍1
🤖 First-ever OCC AI guidance: simple agents can be riskier than sophisticated models
The U.S. banking regulator says examiners should judge AI by the consequences of failure, not model sophistication. A simple AI agent authorized to initiate financial transactions could be riskier than a far more advanced model used only internally.
Higher-risk uses may require tighter limits on autonomy and direct human oversight, per Reuters.
Singularity
The U.S. banking regulator says examiners should judge AI by the consequences of failure, not model sophistication. A simple AI agent authorized to initiate financial transactions could be riskier than a far more advanced model used only internally.
Higher-risk uses may require tighter limits on autonomy and direct human oversight, per Reuters.
Singularity
😁1
🤖 Anthropic open-sourced FlashProve, a system that cryptographically proves which AI model produced a given output.
It works like an email signature: labs publish a "fingerprint" for each model that doesn't reveal weights, so users can verify an answer really came from the claimed model — even against the lab itself. That would block secretly swapping in a cheaper model or inflating results with a pricier one.
Researchers released working code on GitHub. It's intermediate research, not a deployed product — "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it."
Singularity
It works like an email signature: labs publish a "fingerprint" for each model that doesn't reveal weights, so users can verify an answer really came from the claimed model — even against the lab itself. That would block secretly swapping in a cheaper model or inflating results with a pricier one.
Researchers released working code on GitHub. It's intermediate research, not a deployed product — "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it."
Singularity
🥰1