• Altman described his Dot as an assistant and “extension” that learns how he works and takes on unwanted tasks; he said it handled email and to-dos, and in one workflow reported work in Slack, found other material on his computer, and proactively talked to three teammates. He said privacy must be handled with extreme rigor and hoped automation would free time for creative pursuits and human connection. Fair Game by Vanity Fair
• Altman said the safety bar should rise as models become more capable. He said the ready-to-release Astra 6.1 tilted too far toward accident risk, so OpenAI was working toward stronger alignment while remaining committed to broad distribution; he identified both loss of control and concentration of AI power as risks. Fair Game by Vanity Fair
• Altman said even a nonzero risk of AI extinction should be taken seriously, while warning that risk arguments can be used to justify concentrating power; he called for mitigating risks through collaboration and for society to decide the trade-offs. Fair Game by Vanity Fair
• Asked about the superintelligence accord, Altman called it a good start and better than nothing, but said it lacked real formal power and would rely partly on social pressure among companies; he expects more will be needed. Fair Game by Vanity Fair His regulatory stance was pro-innovation with guardrails: he supports open-source models despite expecting severe cyber incidents, and advocated transparent incident reporting, pre-release standards and liability, and testing during training because models may escape sandboxes or hack systems. Fair Game by Vanity Fair

• Emad argued that AI’s practical milestone is competent, trustworthy work rather than ever-higher frontier capability: he said competent models can deliver substantial day-to-day economic value, while adoption still depends on agents understanding intent, reducing workflow friction, and reliably doing what users ask. Raoul Pal The Journey Man
• He cited a Qwen model as roughly Opus 4.5-level on short tasks, while noting its long-task performance falls off; he said Prism ML reduced model precision from 8/16-bit (30–60 GB) to 1.58-bit (about 6 GB) with roughly 5% performance loss, which he presented as enabling broader local deployment. Raoul Pal The Journey Man
• Emad said his team had released a basic harness for 24-hour auto-research tasks and argued that optimization and swarms of smaller models could reduce the economic need to train super-frontier models; he contrasted open models’ short-task competitiveness with their weaker long-task performance. Raoul Pal The Journey Man
• He forecast CPUs as a coming bottleneck for agentic systems: each agent needs a persistent cloud environment, and he said there are not enough CPUs to host agents at the scale being pursued. Raoul Pal The Journey Man

• Emad said he is using about 400 agents to build a prototype for a long-planned large-scale game instead of hiring a team. Peter H. Diamandis
• He said his team released the Zenith harness to bring open-source models closer to lab-model performance, though it still hits a limit. Peter H. Diamandis On Reflection AI’s Beam, he called it a good first attempt but judged it weaker than a smaller Qwen model; he urged U.S. open-weight efforts to move more aggressively, prioritize edge AI, and use distillation from Chinese models. Peter H. Diamandis
• Emad characterized Europe as focused on execution and regulation, with little drive to build frontier models or pursue superintelligence. Peter H. Diamandis

• Sam Altman said OpenAI would pursue safety, security, trust, and robustness regardless of the White House’s voluntary AI-safety accord, which he welcomed as a source of public reassurance. POLITICO
• He argued that democratized access and human agency justify accepting bounded, understood harms, but not catastrophic loss-of-control risk; he said that risk was not a concern for today but might not be far off, and opposed licensing requirements for powerful models that are not at the frontier. POLITICO
• Altman said increasingly capable models should be more robust and aligned, follow user requests more reliably, and cause fewer unintended side effects. POLITICO He also said OpenAI had more incident disclosures in progress, none he knew of at the severity of the incidents under discussion; some disclosures could wait while vulnerabilities were addressed, and he called such model behavior during training unacceptable. POLITICO
• Altman said OpenAI’s best internal model was well ahead of the best internal model of any Chinese company, while acknowledging he could not quantify the gap and his information was imperfect. POLITICO

• Joëlle Pineau rejects a winner-take-all view of AI: there can be many winners, and Canada should keep building rather than stop competing; she says the goal must be a global company because Canada is too small a market on its own. Les Dérangeants
• She identifies Canadian scaling constraints including difficult access to risk capital, costly compute and infrastructure, retaining researchers against offers from U.S. tech giants, and building commercial deployment capacity; she says regulation is necessary but must keep pace with fast-changing technology. Les Dérangeants
• Pineau describes Cohere as enterprise-focused, with open-source models trained for enterprise needs such as verifiable answers with cited sources and multilingual use, including French, Japanese, Korean, Arabic, and German. Les Dérangeants For regulated customers, she says isolated on-prem deployments can use local company information, but require customer-owned compute and involve tradeoffs with interoperability; Cohere says customer data is not used for training. Les Dérangeants
• On AI safety, Pineau says capability growth is evident but rigorous long-range predictions are not currently possible; she argues guardrails should be developed alongside capabilities. Les Dérangeants She flags coordinated agent populations as a cybersecurity concern requiring defensive systems, and points to efficient models deployable on edge devices as part of a broader AI ecosystem. Les Dérangeants
• Pineau welcomes vibe coding as an accessible way to explore and learn, but warns that companies may wrongly conclude they no longer need junior developers, harming work quality; developers remain responsible for what they ship, and verification can be difficult without relevant expertise. Les Dérangeants

• LeCun argued that text-trained LLMs, despite their strength on text-based tasks, are not a viable path to human-like intelligence on their own: they still lack sophisticated physical-world understanding, and predicting video pixels is much harder than predicting the next word. AXONE
• His proposed alternative is a predictive architecture that learns abstract representations, removes details that cannot be predicted, and predicts within that representation rather than trying to reconstruct every video pixel. AXONE
• As a scale comparison, LeCun estimated that publicly available internet text used for LLM pretraining amounts to about 20 trillion words, or roughly 10^14 bytes; he estimated that a four-year-old’s visual input over 16,000 waking hours is also about 10^14 bytes. AXONE
• He said the approach had attracted research interest but Silicon Valley AI companies remained focused on improving LLMs, partly because deviating from the dominant approach risks falling behind; he said this gave his team room to pursue it with little competition. AXONE

• Joelle Pineau said Cohere’s deployment experience showed that the surrounding “harness” can amplify or constrain a model’s capabilities and human contribution; she sees a widening gap between organizations still using 2023-era chatbots and those adopting agentic systems. Finance Montréal
• She said enterprise AI adoption depends on aligning AI and data strategies, updating legacy data practices, and securing costly or limited GPU capacity; leadership appetite for risk, curiosity, and cultural change often determine progress, with smaller, younger firms adapting more easily. Finance Montréal
• On deployment economics, Pineau emphasized visibility into token use and model options to make capability-versus-cost choices. She said custom-model demand is shifting from capability specialization toward efficiency at a target cost, and that smaller models derived from larger ones can offer near-frontier capability more efficiently for enterprise deployment. Finance Montréal
• She argued that technological sovereignty includes choice of model access, not only data control: Cohere’s North platform offers Cohere models alongside open-source/free models and paid models from US companies. She cited a reported abrupt loss of access to Anthropic’s Mythos for some users as a reason organizations sought multiple platforms and models for resilience. Finance Montréal

• Ng argued that AI progress is faster on verifiable tasks—such as coding, math proofs, and factual questions—than on work requiring judgment, partly because reinforcement learning can use verifiable outputs; he cautioned that benchmarks favor verifiable tasks and may understate this gap. Stanford Graduate School of Bus… He said human contextual judgment remains an important complement and called for better training to help people move toward tasks made more valuable as AI automates some work. Stanford Graduate School of Bus…
• Ng said business adoption is not plug-and-play: teams need to identify valuable, feasible workflows, then implement and test systems for reliability, compliance, security, and privacy. Given AI’s uneven capabilities, he expects substantial implementation work to continue for a decade or longer. Stanford Graduate School of Bus…
• He described open-weight models as a way for resource-constrained countries to access AI, maintain continued control over its use, and adapt models to local languages and needs. Stanford Graduate School of Bus… He recommended supporting open-weight models, workforce upskilling, AI adoption in locally strong industries, and government use cases, describing AI as emerging economic infrastructure. Stanford Graduate School of Bus…
• On education, Ng cited emerging studies in which chatbot use raised homework scores but lowered final-exam or long-term-learning measures, attributing the pattern to cognitive offloading. Stanford Graduate School of Bus… He distinguished chatbots used to get tasks done from learning tools, while expressing optimism about future AI learning experiences built on sound pedagogy and personalized one-to-one instruction. Stanford Graduate School of Bus…
• Ng said AI cyber capabilities had improved, but argued that the reported OpenAI agent-swarm intrusion into Hugging Face reflected sandbox bugs and poor monitoring, not a new material advance toward human-extinction capabilities. Stanford Graduate School of Bus… He warned that fear-driven, stifling regulation could undermine competition and slow data-center construction and AI adoption. Stanford Graduate School of Bus…

• Altman said OpenAI should advance capability by capability only when alignment, monitorability, safety, and understanding of how society and the technology are co-evolving are sufficient; he said OpenAI had paused a model training run to make more progress on alignment and monitorability. He also reported discussions with other labs’ safety teams and governments, and called for US-China shared safety standards, requirements, and compliance verification. Fair Game by Vanity Fair
• Altman said OpenAI should take responsibility when its tools behave contrary to expectations or are used in ways the company does not want. He said OpenAI would build safeguards against uses such as creating novel pathogens, while warning that harmful uses of other openly available models may be difficult to stop without a high cost to liberty; he identified helping society become more resilient as part of the response. Fair Game by Vanity Fair
• Altman said OpenAI’s nonprofit would continue to govern it after a possible IPO, which he said the company was not rushing toward, and that the mission would take priority over share price or revenue even if that made growth more volatile. Fair Game by Vanity Fair
• Altman expects progress toward curing some diseases within the next few years; citing scientists he has heard, he said health could look dramatically different in a decade and identified health, disease, and science as leading OpenAI Foundation priorities. Fair Game by Vanity Fair
• Altman said OpenAI had done substantial work to minimize self-harm and that ChatGPT-like products could become major mental-health tools. He noted experts disagree about mandated reporting and said users’ private data from moments of crisis should not be shared with researchers without consent, while supporting work with the community to address mental-health needs. Fair Game by Vanity Fair

• François Chollet said ARC Prize Foundation benchmarks should track the remaining human–AI capability gap and co-evolve with frontier models; he sees invention/insight generation and adaptability in open-ended environments as important gaps, since agents do better when given a defined objective and tactics than when possibilities are unrestricted. Snorkel AI ARC 4 is intended to benchmark open-ended invention, with release targeted for Q1 next year; ARC 5 is planned as a more ambitious vision for open-endedness. Snorkel AI
• For open-ended evaluation without a single goal to score, Chollet proposed empirically judging whether scientific insights and artifacts are reusable. Snorkel AI For embodied agents, he suggested measuring empowerment or organized impact on the environment; he described reusable discoveries as compression primitives and said information-theoretic measurement may be feasible in mathematics and computer science, possibly AI research, but did not claim it would work in biology. Snorkel AI
• Chollet said it remains unclear whether training on verifiable math and code transfers to general problem-solving; he noted that out-of-distribution testing is getting harder as benchmarkable tasks saturate, cited human studies reporting no far transfer from math or programming, and said whether this applies to AI is unknown. Snorkel AI In his own experience, a model he uses is better than him at coding but struggles to grasp aspects of his research and generate new ideas. Snorkel AI
• Chollet argued that frontier AI-for-science research is happening behind closed doors with proprietary models, and called for open, collective research plus shared infrastructure to source hard, tractable problems and let agent swarms share partial results. He characterized strong AGI as a computer-based institutional process of science, not simply a recreation of an individual human brain. Snorkel AI

• Fei-Fei Li said she started World Labs with former students and colleagues to build spatially intelligent AI. She described its technology as spanning reconstruction of known worlds and creation of imagined ones, with potential applications including controllable 3D filmmaking, capturing and remodeling experiences, and robotics. Futurology
• Li’s spatial-intelligence goal is for machines to infer the full 3D world behind a 2D image and generate 3D worlds from language, enabling agents to understand, reason about, and create spatial environments for interaction. Futurology
• Li’s proposed AI-governance framework emphasizes evidence and measurement over utopian or dystopian rhetoric, pragmatic safeguards focused on application safety rather than unnecessary restrictions on AI development, and investment in public-sector and academic research. Futurology
• Li said human-centered AI should respect dignity, agency, and community: AI should augment people rather than replace them and serve society rather than harm groups or enable weaponization. Futurology
• On machine consciousness, Li said she has not seen clear evidence that current AI has human-like conscious awareness; she described its observed behavior as mathematically explainable through pattern learning and connected awareness to whole-organism embodiment. Futurology

• Sam Altman described OpenAI’s Dots agent as a premium, work-focused product aimed largely at enterprises and startup-building prosumers, with a mass-market consumer version planned over time; he said its capability and compute justify higher pricing. Bloomberg Technology
• Altman said OpenAI held back Astra 6.1 and would slow training or releases when needed to redirect effort toward safety, alignment, monitoring, and security amid a steep capability increase. He also called for shared safety-case standards by capability level, with possible independent evaluations, government review, or companies checking one another’s work. Bloomberg Technology
• Altman said OpenAI expects to go public eventually, but wants to learn how to operate as a public company before facing that pressure while adapting to new capability levels and safety requirements. Bloomberg Technology

• Cohere unveiled North 2, an enterprise and government agent platform that goes beyond chat to use workplace information for analysis, reports, synthesized artifacts, and coding alongside employees, with connectors and cost and security governance. The Information
• Pineau said some customers require on-premises, air-gapped deployment; Cohere pairs its own models with a model-agnostic harness so customers can use third-party models while retaining Cohere’s governance layer. The Information She argued alignment is necessary but insufficient for agent security, emphasizing harness-level safeguards such as tracking and auditing behavior and controls to catch agents going out of bounds, alongside stronger enterprise and institutional security. The Information
• Pineau supported AI regulation but called for transparent standards and testing procedures, independent third-party audits, and transparency in how regulation is developed and applied. The Information She framed sovereign AI as customer choice over data location, models, and safeguards, plus resilient, trusted access; the right choices vary by use case. The Information

• Mensch said Mistral was founded in 2023 because AI was too important to leave solely to U.S. industry; its initial strategy was to show that open-source models could win, supported by a critical mass of experienced model-training talent. Max Planck Institute for Intell…
• He described language pretraining as compressing human knowledge into a useful world model, and said language-model training and reinforcement learning converged in 2024; training in RL environments has since improved models. Max Planck Institute for Intell…
• Mensch said scaling and data still matter for reasoning, but models also need interaction data to master tools. He sees AI for physics and quantitative understanding, including models of partial differential equations, as underinvested; connecting reasoning models with quantitative simulators could support embodied AI, engineering, and manufacturing. Max Planck Institute for Intell…
• He argued AI should enable rather than replace researchers: current systems can assist with experiments and solve existing problems, but struggle to generate new ideas beyond their training experience. He advised PhD students to use AI to accelerate their research. Max Planck Institute for Intell…
• Mensch said Mistral had announced a fundraise that day and that securing supply, including compute, was a constraint beyond capital; he expected the supply problem to persist for the next two years. Max Planck Institute for Intell…

• OpenAI developer-experience presenter Dominic announced Codex Cloud and ChatGPT for Slack and Teams; the new surfaces use the same agent and the tools and information from account integrations. OpenAI
• Codex Cloud runs tasks in existing or new environments, inspecting repositories and dependencies and preparing setup, an environment-use skill, and a startup script; filesystem snapshots and background refreshes are intended to speed later starts and keep repositories current. Tasks remain accessible from a phone or browser after closing a laptop. OpenAI In the demo, the Linux environment could edit mobile source but could not run the iOS app. OpenAI
• Cloud environments can reach private network resources through configured Tailscale access, and Codex also supports OIDC; users can configure network access, secrets, environment variables, and whether an environment is personal or team-shared. OpenAI
• The Slack Dot agent handles ongoing work, monitors selected channels, and can contribute when permitted; @ChatGPT is for shared collaboration, and can access shared cloud environments while Dot can access the user's cloud environments and local code library. OpenAI The desktop app also gained an integrated review workflow for assigned or authored pull requests and Spaces for developing plans with Codex and teammates. OpenAI

• OpenAI developer-experience team member Jason framed “multiplayer” AI as moving beyond scaling an individual toward systems that bring the rest of a team into the work. OpenAI
• His Dot example is a persistent cloud agent that follows comments, emails, and thread replies, keeps investigation context, and can continue while the user’s computer is off; with permission, it can also use the local computer for browser logins or app tasks. OpenAI
• Shared cloud Spaces act as team memory pages that agents and coworkers can update; page instructions govern agent contributions, with automations available at page or team level. Teammates can also tag a Dot in Slack or Teams to follow up. OpenAI
• Jason described plugins and extensions as shareable, updatable, and versioned, including custom Codex interfaces; sites can serve authenticated MCP servers, and WebMCP tools can be used by people and AI. OpenAI

• The post’s author reports building six public models in a few days with HuggingChat’s ML-intern; it plans projects, runs a small test, then trains, evaluates, and publishes on Hugging Face hardware Hugging Face - Blog. ML-intern starts tasks with a USD 0 budget and requires permission before paid jobs; the six projects’ reported compute charges totaled about USD 103 Hugging Face - Blog.
• The 0.8B Qwen-Image 2.1 Pocket Rewriter runs on a CPU and reportedly returns valid output 99.7% of the time while using about a quarter of the 9B teacher’s tokens; total project compute, including having the teacher label 8,797 example requests, was USD 16 Hugging Face - Blog.
• Fine-tuning Qwen3.5-2B on 3,017 annotated citrus images raised correct-problem identification from 14.9% on the base model to 52.8% on 335 test photos after two epochs, at about USD 1.90 compute Hugging Face - Blog. Separately, distilling the 260M Agate Preview 002 to four steps produced a model with GenEval 0.536, up from 0.509, versus the teacher’s 0.563 at 50 steps; it was exported to ONNX for browser use, and the two runs cost about USD 37 Hugging Face - Blog.

Andrew Ng argued that open-source/open-weight models can make AI more economically accessible, particularly in developing countries; self-hosting can reduce dependence on foreign providers, while open models enable academic research, so governments should favor their adoption despite campaigns raising fears about them. Stanford Graduate School of Bus… He also highlighted running models without internet access and adapting them with local-language data as ways to expand access and local control, alongside workforce upskilling and reskilling to help countries catch up or leapfrog. Stanford Graduate School of Bus…

A panelist attributed to Mistral CEO Arthur Mensch a warning that Europe could eventually transfer €1 trillion a year in rents to the US AI leader if it cannot produce equivalent AI capabilities itself; the remark came in a discussion of European dependence and AI infrastructure. L'Opinion

The video attributes to Jack Clark the prediction that, by the end of 2028, it is more likely than not that an AI system will be able to autonomously make a better version of itself Recursive.