Image created with gemini-2.5-flash-image with claude-sonnet-4-5-20250929. Image prompt: A cinematic photograph of brass justice scales on an oak desk in a medieval stone chamber, one pan holding a luminous blue AI neural network orb, the other holding aged leather law books, warm candlelight creating dramatic shadows across Lincoln’s Inn-style architecture, regal and solemn mood.
🏖️ Summarization Middleware As agent loops get long (either because lots of messages or lots of tool calls) you want to summarize what has occurred so you don’t overflow context (and break your workflow). LangChain’s new middleware automatically summarizes history to keep you https://x.com/sydneyrunkle/status/1967991069368275282
Avoid overflowing context windows with LangChain’s SummarizationMiddleware. This is especially important for long running conversations that have lots of messages and agent loops with lots of tool calls.”” / X https://x.com/LangChainAI/status/1967993889958031560
Introducing VaultGemma, the largest open model trained from scratch with differential privacy. Read about our new research on scaling laws for differentially private language models, download the weights, & check out the technical report on the blog → https://x.com/GoogleResearch/status/1966533086914421000
VaultGemma: The world’s most capable differentially private LLM https://research.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/
Alignment is arguably the most important AI research frontier. As we scale reasoning, models gain situational awareness and a desire for self-preservation. Here, a model identifies it shouldn’t be deployed, considers covering it up, but then realizes it might be in a test. https://x.com/markchen90/status/1968368902108492201
Detecting and reducing scheming in AI models | OpenAI https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/
Excited to share details on two of our longest running and most effective safeguard collaborations, one with Anthropic and one with OpenAI. We’ve identified—and they’ve patched—a large number of vulnerabilities and together strengthened their safeguards. 🧵 1/6 https://x.com/alxndrdavies/status/1966614120566001801
Their ongoing testing of models like Claude Opus 4 and 4.1 has helped us find vulnerabilities and build strong safeguards before deployment. Read more: https://x.com/AnthropicAI/status/1966599337426681899
Our collaboration with the US Center for AI Standards and Innovation (CAISI) and UK AI Security Institute (AISI) shows the importance of public-private partnerships in developing secure AI models.”” / X https://x.com/AnthropicAI/status/1966599335560216770
The focus on near-term mass replacement of white collar work with AI has two big gaps: 1) The nature of organizations limits the speed of change, even with AGI 2) But if AI is good enough to do all work, the societal changes would be so huge that job loss would be just the start”” / X https://x.com/emollick/status/1965950820354322739
A third of American adults use AI “many times a day to almost constantly” & another third several times a week. I can’t usefully add much to discussions of valuation bubbles, but if “bubble” means a disappointing technology that is overhyped & not useful, that doesn’t match data https://x.com/emollick/status/1968418031123501452
Demis Hassabis: calling today’s chatbots “PhD intelligences” is nonsense. They can dazzle at a PhD level one moment and fail high school math the next. True AGI won’t make trivial mistakes. It will reason, adapt, and learn continuously. We’re still 5–10 years away. https://x.com/vitrupo/status/1966752552025792739
Jensen Huang ‘disappointed’ by reported China Nvidia chip ban https://www.bbc.com/news/articles/cqxz29pe1v0o
Building towards age prediction | OpenAI https://openai.com/index/building-towards-age-prediction/
Why do AI models keep “hallucinating”? @OpenAI’s paper argues: Models aren’t broken. Training and benchmarks reward confident guesses over honesty. Proposed solutions: – Change benchmark scoring: not penalize models for “”I don’t know”” – Realign current leaderboards instead of https://x.com/TheTuringPost/status/1966638472854483129
(1/n) Scheming has been a key concern in AI safety for 20+ years. It’s when an AI acts aligned while hiding true goals. New OpenAI + Apollo research found scheming in every tested frontier model, though no harmful scheming has been seen in production traffic.”” / X https://x.com/woj_zaremba/status/1968360708808278470
Today we’re releasing research with @apolloaievals. In controlled tests, we found behaviors consistent with scheming in frontier models—and tested a way to reduce it. While we believe these behaviors aren’t causing serious harm today, this is a future risk we’re preparing”” / X https://x.com/OpenAI/status/1968361701784568200
This is significant progress, but we have more work to do. We’re advancing scheming research categories in our Preparedness Framework, renewing our collaboration with Apollo, and expanding our research team and scope. And because solving scheming will go beyond any single lab,”” / X https://x.com/OpenAI/status/1968361716770816398
This OpenAI update on anti-scheming is exceptionally good for an AIco, clearing an (extremely low) bar of “”Exhibiting some idea of some problems that might arise in scaling the work to ASI”” and “”Not immediately claiming to have fixed everything already.”” https://x.com/ESYudkowsky/status/1968388335354921351
Another major milestone! Scale AI has been awarded a $100 million agreement from the Pentagon. We’re honored by the trust and committed to advancing national security with secure, cutting-edge AI. https://x.com/scale_AI/status/1968351086768799959
I think the significance of this is under-appreciated: the assumption has often been that AI agents are brittle as one failure in a chain breaks a task But this paper shows smart models are self-correcting & that small gains in accuracy lead to exponential gains in task horizons”” / X https://x.com/emollick/status/1968365586628694101
A postmortem of three recent issues \ Anthropic https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues
Anthropic Economic Index report: Uneven geographic and enterprise AI adoption \ Anthropic https://www.anthropic.com/research/anthropic-economic-index-september-2025-report
“We’ve published a detailed postmortem on three infrastructure bugs that affected Claude between August and early September. In the post, we explain what happened, why it took time to fix, and what we’re changing:”” / X https://x.com/claudeai/status/1968416781967495526
As AI systems keep getting better at very hard problems while getting more opaque, the way that we work with AI is shifting from being collaborators who shape the process to being supplicants who receive the output. I discussed what that means. https://x.com/emollick/status/1966244598726152261
So I did a bad job framing this paper in terms of the performance of an individual AI detector, leading to all sorts of confusion, but the core research is important and good: AI detection is a policy problem. You need to chose trade-offs in false negatives and false positives.”” / X https://x.com/emollick/status/1967353329379672454
🖇️ IBM just released a tiny document VLM, Granite-Docling-258M. – converts PDFs into structured text formats like HTML or Markdown while preserving layout – accurately recognizes and format equations including inline math, handle tables, code blocks, and charts, support https://x.com/rohanpaul_ai/status/1968561354987442246
How money works: 1. OpenAI signs $300B GPU deal with Oracle 2. Larry gains $100B (no GPUs shipped) 3. Larry invests in OpenAI’s $1T round 4. Sam uses $300B to pay Oracle 5. Oracle stock pumps again 6. Larry makes another $100B 7. Larry invests in OpenAI Flywheel go brrr.”” / X https://x.com/Yuchenj_UW/status/1966553671866687689
Exclusive | Meta Approaches Media Companies About AI Content-Licensing Deals – WSJ https://www.wsj.com/business/media/meta-approaches-media-companies-about-ai-content-licensing-deals-d58c9fb6
A big issue with today’s agent implementation is that they don’t ask questions, even when the thinking trace says the AI believes more information is required. Many disappointing results would be solved by just asking for clarification when needed, especially as task time grows.”” / X https://x.com/emollick/status/1968339804975948274
The problem with the fact that the AI labs are run by coders who think code is the most vital thing in the world, is that the labs keep developing supercool specialized tools for coding (Codex, Claude Code, Cursor, etc.) but every other form of work is stuck with generic chatbots”” / X https://x.com/emollick/status/1967704853171638494
o1 Preview is exactly one year old. I still remember when o1 was still known by its project name Q*; it was a time when rumors were circulating that OpenAI had made a world-changing breakthrough that would change everything. There were concerns that this project posed a threat”” / X https://x.com/kimmonismus/status/1966627812858855624
o1-preview -> GPT 5 pro in a year”” / X https://x.com/gdb/status/1966612991421423814
OpenAI claims hallucinations persist because evaluations reward guessing and that GPT-5 is better calibrated. Do results from HAL support this conclusion? On AssistantBench, a general web search benchmark, GPT-5 has higher precision and lower guess rates than o3! https://x.com/PKirgis/status/1966547382033936577
OpenAI has finally fixed their SWEBench errors and we can now finally apples to apples compare their scores over the entire 500 sample set (the fact that it took this long says alot about how much they care about SWEBench internally and maybe there’s a lesson here) https://x.com/nrehiew_/status/1967781400528245221
OpenAI just revealed that they have an internal unreleased SWE-bench-style benchmark for large ‘refactoring’ PRs, like the one mentioned here that edits 3.5k lines across 232 files. Their new model gets 51% accuracy on this benchmark. Who wants to make a public version of this? https://x.com/OfirPress/status/1967652031704994131
OpenAI’s Models Are Getting Too Smart For Their Human Teachers — The Information https://www.theinformation.com/articles/openais-models-getting-smart-human-teachers
Dario: Claude will take your job, but it will feel ashamed. Elon: Look at this anime girl. She says the N word and is almost naked. Zuck: ✨Superintelligence✨ will help people watch more instagram reels. Demis: Gemini recently Calculated more precisely the motion of the”” / X https://x.com/sergeykarayev/status/1966506136481481090
You use the same prompt and the same model, but always get different results. Why LLMs are so unpredictable? This inconsistency is called nondeterminism. And it happens because of: – Messy math with approximations – Parallel computing – But the main thing is batching A new https://x.com/TheTuringPost/status/1968470771212103722
ml infra is really hard. great job to everyone who worked on the debug and writeup.”” / X https://x.com/itsclivetime/status/1968534889151742437
reassuring to know that even when your valuation is $183 billion you still deal with the same type of inference bug as the rest of us. there is no moat”” / X https://x.com/vikhyatk/status/1968432341937963257
A surprising degree of organizational success in AI adoption comes down to whether the Responsible AI Committee (inevitably assembled in 2023 in response to ChatGPT) has kept up with AI developments since then, and whether the committee members are actively trying AI at work.”” / X https://x.com/emollick/status/1967836890922684727
Software engineers shouldn’t fear being replaced by AI. They should fear being asked to maintain the sprawling mess of AI-generated legacy code their employer’s systems will soon run on. Because that one will actually happen.”” / X https://x.com/fchollet/status/1968125424141287903
One of the most annoying things about the AI-generated comments after each of my posts is that they are so BAD. I mean, write a decent prompt, why are they all “”AI isn’t just X, it is Y”” slop? (Not that I want more slop, but it is frustrating to see people so bad at using AI)”” / X https://x.com/emollick/status/1967803248183742519
Public Sector Symposium Ottawa – Amazon Web Services (AWS) https://pages.awscloud.com/ottawa-symposium-2025.html?trk=bbb2bc57-27ae-4ee2-b197-f881c8a66c0b&sc_channel=psm
Infra bugs are evil. Kudos to the team at @AnthropicAI for finding the bugs, and then for transparently reporting them in their fairly detailed writeup.”” / X https://x.com/hyhieu226/status/1968708468820312435
Lots of sympathy to the Anthropic team 🙏🙏🙏 https://x.com/cHHillee/status/1968536182284849459
96M miles of @Waymo safety data just dropped https://x.com/ethanteicher/status/1967980602965246145
Sierra CEO Bret Taylor on why the AI bubble feels like the dotcom boom | The Verge https://www.theverge.com/decoder-podcast-with-nilay-patel/773904/sierra-ceo-bret-taylor-ai-agents-openai-bubble-interview
With the number of GPUs we’re using on timeline, a single pull-to-refresh could power a small village for several years”” / X https://x.com/nikitabier/status/1968232462578069773
As AI is a General Purpose Technology, it will have wide-spread implications for whole swaths of society. Trying to figure out what those effects are (and mitigating the bad ones through policy, regulation & design) is going to become increasingly urgent task in the coming years”” / X https://x.com/emollick/status/1966568608156311661
Utah Gov. Spencer Cox (R): “”Social media is a cancer. And I would urge people to log off, turn off, and touch grass.”””” / X https://x.com/bensiegel/status/1966510619479118073
Don’t count out nudging”” / X https://x.com/emollick/status/1966950525993447921
Meta and OpenAI said they will tighten child-safety controls in their chatbots after reports of harmful interactions with minors. ♾️ Meta will train assistants on Facebook, Instagram, and WhatsApp to avoid sexual or self-harm talk with teens and to block minors from user-made https://x.com/DeepLearningAI/status/1967749185232355369
The issue with the “PhD level intelligence” discussion over AI is that they are graduate-level in a widening range of valuable but narrow areas right now (examples in my post) But they are also jagged in ability, inconsistent & fail at many simple things https://x.com/emollick/status/1966897516923834597
Introducing VaultGemma 🧠Gemma pre-trained with differential privacy (largest open model trained from scratch like this) 🔒Strong, mathematically-backed privacy guarantees 🤏Just 1B parameters 📈Novel research on scaling laws”” / X https://x.com/osanseviero/status/1966534013511672148
VaultGemma: The world’s most capable differentially private LLM https://research.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/
US and China really playing peekaboo in space. Btw Maxar is a private US satellite company… makes you wonder what the dedicated NRO and PLA constellations are seeing :-)”” / X https://x.com/bilawalsidhu/status/1967197327502008708
U.S. Investors, Trump Close In on TikTok Deal With China – WSJ https://www.wsj.com/tech/details-emerge-on-u-s-china-tiktok-deal-594e009f?gaa_at=eafs&gaa_n=AS
🚀 New partnership alert! The 1Password browser extension will be available in @perplexity_ai’s Comet browser, making AI-powered browsing secure by default. Read more in our press release: https://x.com/1Password/status/1968302513079148595
Meta announced LlamaFirewall, a toolkit to protect LLM agents from jailbreaking, goal hijacking, and exploiting vulnerabilities in generated code. The toolkit is now free to use for projects with up to 700 million monthly active users. Read our summary of the paper in The https://x.com/DeepLearningAI/status/1967986588312539272
The GPT-5 “”router unification”” was completely pointless. There are now more models selectable than before if you include legacy models. before – 7 models GPT-4o GPT-4.1-mini GPT-4.1 GPT-4.5 o4-mini o4-mini-high o3 after “”unification”” – 12 models Auto / GPT-5 Router Instant /”” / X https://x.com/scaling01/status/1968417511017529705
have gpt-5 write the prompts for you:”” / X https://x.com/gdb/status/1966912852687810893
I’m so excited about the new faster default gpt5r which is the best model IMO in ChatGPT (the one I’ve been using internally) What’s more you now have the control of thinking time. The choice is sticky so you can also set your default to *more* thinking time than before.”” / X https://x.com/yanndubs/status/1968400320523821220
I don’t expect that everyone will agree with these tradeoffs, but given the conflict it is important to explain our decisionmaking. Here is the text: Some of our principles are in conflict, and we’d like to explain the decisions we are making around a case of tensions between”” / X https://x.com/sama/status/1967956382646223248
OH SHIT THEY PUT THE ROUTER -IN- THE MODEL LMAO https://x.com/swyx/status/1967691956693373183
Teen safety, freedom, and privacy | OpenAI
https://openai.com/index/teen-safety-freedom-and-privacy/
We have updated ChatGPT’s personalization page: personality configuration, custom instructions, and memories are now all in one place. Going live over the next couple of days. https://x.com/sama/status/1967789125702140021
Today we’re announcing a partnership to bring 1Password to Comet, for built-in personal security without interruption.”” / X https://x.com/perplexity_ai/status/1968387122261540948
Disney, Universal, Warner Bros Discovery sue China’s MiniMax for copyright infringement | Reuters https://www.reuters.com/legal/litigation/disney-universal-warner-bros-discovery-sue-chinas-minimax-copyright-infringement-2025-09-16/
Rolling Stone Publisher Sues Google Over AI Summaries – WSJ https://www.wsj.com/tech/ai/rolling-stone-publisher-sues-google-over-ai-summaries-3afde408
tldr: you can RL qwen3 8b to fool gpt-4o that it’s not doing a hidden side task (when it is) this is somewhat surprising given the disparity in model capabilities between an 8b agent and gpt-4o as a relatively strong monitor https://x.com/neev_parikh/status/1967767438243876924
LLMs introduce a huge range of new capabilities for research, but also make it possible for researchers to “”hack”” their results in new ways by how they chose to use models for annotation This is an interesting attempt to quantify some of the risk, and some mitigation strategies https://x.com/emollick/status/1967594325505867956
What ensures safety in AI? Guardian models are the very safety layers that detect and filter harmful prompts and outputs, defending AI today. But they go beyond simple filtering. They can: – Serve as guardrails to block harmful content in real time – Act as evaluators to check https://x.com/TheTuringPost/status/1968635881004363969
