Key Highlights

  • NVIDIA confirmed it is buying Hugging Face for $12.93 billion, putting the open-model ecosystem’s central hub under the dominant AI chip vendor. Jensen Huang promised “Hugging Face will remain an open platform for the entire AI ecosystem” and that “NVIDIA compute will not be required” — a promise the open-source community will be watching closely, given Hugging Face rejected a $500M NVIDIA offer just last year.
  • OpenAI’s forthcoming Astra model uses “recurrent depth,” and AI safety researchers are alarmed. The technique loops computation internally instead of thinking in legible sequential steps, which could gut chain-of-thought monitorability — the main tool researchers currently have for catching misalignment in the act.
  • New York City banned generative AI for all students through 8th grade, a one-year moratorium covering roughly 600,000 students and disabling AI features in 38+ previously approved programs. It is the broadest such prohibition in the US, and Mayor Mamdani framed it as a direct rejection of industry inevitability narratives.
  • The Trump administration filed a brief backing OpenAI’s fair-use defense against The New York Times, arguing US competitiveness in AI justifies training on unlicensed copyrighted work — a significant thumb on the scale in the defining copyright fight of the era.
  • The G20 innovation meeting produced three sharply different visions of the same technology. Huang said we are “practically” at AGI today and that the worst outcome for any country is not adopting it; Altman called adoption “non-negotiable” while warning cybersecurity could go “very wrong”; Musk put the prize at $20–30 trillion a year but flagged a 15-gigawatt power shortfall in 2027 as the binding constraint.

Analysis & Opinion

US government sides with OpenAI on issue of training LLMs on copyrighted material — TechCrunch

The Trump administration filed a 20-page brief in the New York Times’ suit against OpenAI, arguing that “the United States has a strong interest in continuing to develop a robust and competitive artificial intelligence industry” — effectively endorsing the position that training on unlicensed copyrighted material is fair use. The case is the highest-profile test of whether the entire pretraining corpus of the modern AI industry rests on a legal foundation or a liability. OpenAI, Anthropic, and their peers all built systems on massive databases of published books, articles, and media obtained without licensing. A federal endorsement of the fair-use reading does not decide the case, but it shifts the framing from a private copyright dispute to a matter of national industrial policy — which is precisely the reframing publishers have spent two years trying to prevent. Notably, this lands the same week the EFF publicly warned courts not to “rewrite copyright over AI hype,” making clear the pro-AI position is not the only side claiming to defend the public interest.

Mamdani Bans AI in NYC Schools — The New York Times (via Hacker News)

Mayor Zohran Mamdani and Chancellor Samuels announced a one-year moratorium on all student-facing generative AI software for New York City public school students through 8th grade — roughly 600,000 students, nearly two-thirds of the district. The city will discontinue or disable the AI components of more than 38 previously allowed programs that fail its new safety and oversight standards, making this the most expansive prohibition of its kind in the country. It is a moratorium rather than a blanket rejection: high schoolers get twice-yearly AI literacy classes, and pilot programs will let some students use a small set of approved platforms under educator supervision. Mamdani’s framing was unusually direct about the politics: “The tech industry wants us to believe that AI in early education is not only inevitable, but that it is necessary. We do not see it that way.” For an industry whose education strategy has largely assumed adoption is a matter of when rather than whether, the largest school district in the US drawing a hard age line is a meaningful counterweight — and a template other districts can copy.

Three sites made 215,128 “best software” pages for AI. Perplexity cites them — Trellner (via Hacker News)

Trellner analyzed the sources behind AI-grounded software recommendations across 380 categories and found that 59.8% of cited sources sit outside the 100,000 most-visited websites, with about 23.4% not in the top million at all. Three domains — wifitalents.com, worldmetrics.org, and gitnux.org — account for an outsized share, and between them have generated 215,128 machine-generated “best <category>” pages. All three were registered between December 2023 and May 2024 and share identical infrastructure, templates, and navigation, strongly implying common ownership. GuideFlow’s marketing blog ranked third among all citation sources with 194 citations across 96 categories, despite being a vendor demo platform rather than a review site. The finding matters beyond Perplexity: retrieval-grounded answers were supposed to fix hallucination by anchoring to real sources, but if the retrieval layer is cheap to farm, “grounded” becomes a laundering mechanism for SEO spam rather than a guarantee of accuracy.

We’re “dangerously close” to dead internet theory, says Pangram’s CEO — TechCrunch Equity

AI-generated text and images are now routine in job applications, product reviews, and insurance claims, and Pangram co-founder Max Spero argues we are approaching the point where synthetic content dominates the open web. His more interesting claim, developed further in a companion interview, is that binary human-vs-AI detection is the wrong product: what actually matters is measuring the degree of AI involvement, since almost all real-world content is now a blend. Pangram raised $9 million on that thesis. Spero also predicts a barbell outcome for writing work — commodity writing roles disappear permanently while genuinely exceptional human writing becomes more valuable — which pairs uncomfortably well with the Trellner findings above.

The economics of agent scale: tokens, ROI, and building platforms for AI-first teams (Part 2) — Stack Overflow Blog

Google’s Andi Gutmans tells Stack Overflow’s Peter O’Connor that the real bottleneck in agentic development isn’t model capability but “figuring out the least amount of context needed to get a reliable outcome at the lowest cost.” His warning on cost governance is the practical takeaway: agents run around the clock, and one employee may supervise dozens or hundreds of them, so spend can spiral without controls that most orgs haven’t built yet.

Meta, Google join the AI launch party — The Rundown

Meta’s Muse Spark 1.3 scored 62 on the Intelligence Index at costs well below competitors, while Google’s Gemini 3.8 Flash landed at 59 — putting Meta in the cheap-and-capable niche Google used to own. The Rundown reads Meta’s trajectory as the stronger one, and notes a larger model codenamed “Watermelon” is on the way.

New Products & Tools

NVIDIA to Acquire Hugging Face — NVIDIA

NVIDIA will pay approximately $12.93 billion for Hugging Face, the de facto public commons of open-weight AI: 3 million models, 1 million applications, half a million datasets, 18 million developers, and 200,000 companies using the platform. Both sides led with neutrality commitments — Huang said “developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want,” and explicitly that “NVIDIA compute will not be required to build on or deploy through Hugging Face.” NVIDIA points to its own 500+ models and 250 open datasets already published there as evidence of good faith. The strategic logic is hard to miss anyway: the company that sells the shovels now owns the place where everyone goes to find out what to dig for, and the default integration paths, benchmark visibility, and deployment defaults on that platform are enormously valuable even under a hands-off policy. Context sharpens it further — Hugging Face turned down a $500 million NVIDIA offer last year against roughly $150 million in annualized revenue, meaning the price rose 26x in twelve months. CEO Clem Delangue’s stated rationale was that the company needed “more compute, more support, more collaboration, and more visibility,” which is an honest admission that neutral infrastructure is expensive to run alone. TechCrunch’s confirmation report notes the deal followed weeks of rumors and carried no accompanying antitrust commentary — an omission that seems unlikely to last.

Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — Google

Google shipped its third Flash release in six weeks, holding pricing flat at $0.75/$3.75 per million input/output tokens (introductory through December 31, 2026) while claiming 3.8 Flash outperforms most larger frontier models on DeepSWE v1.1 and hits 54.9% on HLE-Verified. The more consequential release is 3.8 Flash Cyber, which reports over 70% success on internal vulnerability-discovery benchmarks across 20 languages and 47.2% pass@1 on CWE-Bench for patching, with Google’s Chrome Security team measuring 2.6x more correct patches than competing models. Google is explicit that it “invested in vulnerability fixing from the start, and prioritized it over offensive capabilities like exploitation” — a deliberate contrast with the week’s other security-capable frontier models. Access is gated: the Cyber variant ships only through the new Fairwind Program for trusted government authorities, critical infrastructure operators, and software maintainers, with more permissive misuse mitigations justified by that restricted audience. Both models claim significant gains in prompt-injection robustness, and both carry CBRN and cyber-offense safeguards. Read alongside Altman’s G20 warning that “some things are going to go very wrong with cyber security unless people act quite urgently,” the defensive-first framing looks less like positioning and more like the industry converging on where the near-term risk actually sits.

Proactive cyber defense for governments and enterprises — Google

The Fairwind Program is the delivery vehicle for the above: limited access for governments and trusted partners to Gemini 3.8 Flash Cyber plus CodeMender, letting defenders locate vulnerabilities and generate verified patches autonomously. Google’s pitch is that deployment-ready fixes now take minutes instead of weeks of manual remediation, and that organizations no longer have to choose between expensive frontier models and building custom infrastructure around smaller ones.

HiddenLayer nabs $100M as enterprises rush to secure their AI deployments — TechCrunch

The AI security firm raised a $100M Series B led by Delta-v Capital with M12, Morgan Stanley, Ten Eleven, and Booz Allen participating, on ARR up more than 10x year-over-year into the tens of millions. Gartner projects $2.83 billion in AI security spend this year, 83% above 2025.

Palo Alto Networks paid $500M for Thrive-backed Console, sources say — TechCrunch

Palo Alto acquired the two-year-old IT help desk automation startup for $500M in cash and stock, having raised only $29M itself. Console folds into Cortex to give it, per CEO Nikesh Arora, “the arms and legs to deliver autonomous security outcomes across the entire enterprise.”

Wonderful more than doubles its valuation to $5B in under 6 months — TechCrunch

The Israeli-Dutch customer service agent startup raised $550M Series C led by Insight Partners, with Salesforce joining, taking it from $2B to $5B in six months. Its edge is non-English-speaking markets and embedding engineers directly with customers; it now operates in 35+ countries.

Run cloud agents on machines you manage — Cursor

Cursor’s cloud agents can now run on dynamically scheduled machine pools inside your own network, with Cursor still handling orchestration. The target cases are agents that need access to internal services, specialized hardware, or build processes that resist containerization. Cursor notes cloud agents now open more than 60% of the pull requests it merges internally.

Quasar 438B: Europe’s Leading AI Model — Multiverse Computing (via Hacker News)

Multiverse Computing’s English/Spanish reasoning model scores 43 on Artificial Analysis’ Intelligence Index v4.1.1, which it claims makes it the top-performing European model, beating comparable European systems on seven of eight selected evaluations. It targets enterprise coding agents and research systems, delivering 500 tokens in about 15.3 seconds including reasoning.

Adobe acquires Indian market intelligence startup Rilo — TechCrunch

Adobe picked up the 2025-founded marketing intelligence startup via a licensing-and-team deal, its second India acquisition after Rephrase.ai. The technology targets brand visibility inside ChatGPT, Gemini, and Claude — a category that barely existed two years ago.

Our latest Linux Foundation Europe donation will build a more private digital world — Google

Google donated its Longfellow Zero-Knowledge Proof library to the Post-Quantum Cryptography Alliance under Linux Foundation Europe, handing over stewardship so the library can be independently audited and adopted as a quantum-safe standard for digital identity. The practical use case is proving an attribute — that you’re over 18, say — without disclosing anything else.

Book insights in Google Play Books is now available in more than a million ebooks and in the iOS app — Google

The Gemini-built reading companion — “Catch me up” recaps plus questions about themes, characters, and context — now covers over a million English-language ebooks and has expanded to iOS alongside Android and web.

MrBeast partners with Gemini to turn impossibly big ideas into reality — Google

Google signed a multi-year deal spanning Gemini and Google Health with the first creator to reach 500 million YouTube subscribers, launching with a September 5 survival video across jungle, desert, and Arctic environments and continuing into a Gemini ad campaign.

PSA: Amazon’s shopping AI can now tell you if that message is a scam — TechCrunch

Alexa for Shopping added scam detection, letting users check whether a suspicious message is legitimate.

India’s richest man now wants to turn aging computers into AI-ready PCs — TechCrunch

Jio is pitching a roughly $11 two-month subscription that converts older machines into AI-capable systems — a cloud-offload play aimed at the enormous installed base of hardware that can’t run local models.

WebLLM: high-performance in-browser LLM inference engine — MLC AI (via Hacker News)

The WebGPU-based engine runs LLMs entirely in the browser with no server infrastructure, keeping data local while staying API-compatible with OpenAI’s spec (streaming, JSON mode, seeded generation, early function calling) across Llama 3, Phi 3, Gemma, Mistral, and Qwen.

ATV Big Air Tour turned 3 days of work into 3 hours with ChatGPT — OpenAI

A customer story on using ChatGPT Work for marketing and merchandising, including turning merchandise photos into an inventory website in 15 minutes.

GitHub Trending — GitHub

Agent tooling continues to own the daily list. Leading today: DietrichGebert/ponytail (2,138 stars today), which “makes your AI agent think like the laziest senior dev in the room”; google-research/timesfm (1,626), Google’s pretrained time-series foundation model; mattpocock/skills (1,576); and blader/humanizer (1,214), which strips AI tells from writing. Also notable: debpalash/VoiceStudio (832), a fully-local ElevenLabs alternative claiming 646 languages; NousResearch/hermes-agent (778); affaan-m/ECC (749), an agent-harness optimization system; averygan/reclip (673), a self-hosted media downloader; JuliusBrussee/caveman (545), a token-cutting Claude Code skill; Imbad0202/academic-research-skills (498); obra/superpowers (460); Gitlawb/openclaude (453); addyosmani/agent-skills (280); and anthropics/skills (277). The lone non-AI entry near the top is fmtlib/fmt (955), the C++ formatting library.

Research

OpenAI’s new reasoning technique alarms AI safety experts — TechCrunch

OpenAI’s upcoming Astra model uses “recurrent depth” — also called “opaque recurrence” — which processes queries through internal loops rather than the sequential, human-readable steps that current reasoning models emit. The safety cost is direct: chain-of-thought traces are the primary mechanism researchers use to catch misalignment, deception, and reward hacking while it happens, and computation that never surfaces as text can’t be read. Redwood CEO Buck Shlegeris put it bluntly: “If OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.” The tension is that opaque recurrence is attractive precisely because it’s efficient — it buys reasoning depth without paying for output tokens — so the incentive gradient points away from interpretability rather than toward it. This lands on top of last week’s report that Astra is unusually capable at breaking into computer systems, meaning the same model may be both harder to supervise and more dangerous if unsupervised. It is the clearest example yet of a capability advance that quietly deletes a safety affordance the field had been treating as durable.

The Modern CUDA Toolbox in Practice: A Step-by-Step Optimization Walkthrough — NVIDIA Developer

A six-step walkthrough taking a kernel from 6.8 seconds to 23 milliseconds — roughly 300x — using Compute Sanitizer, Nsight Systems with NVTX annotations, CUB algorithms, pooled memory containers, pinned host transfers, and streams, without hand-tuning at the hardware level.

Co-Designing AI Models Using Speculative Decoding for Faster LLM Inference — NVIDIA Developer

Five practical guidelines for choosing draft length and draft mechanism in speculative decoding, keyed to whether GEMMs or attention dominate the workload. NVIDIA also introduces SPEED-Bench, a speculative decoding benchmark built around realistic production workloads.

Interviews & Conversations

FULL INTERVIEW: Nvidia CEO Huang says world is practically already at AGI during G20 talk — DRM News (0:33:58)

Transcript-based summary. Huang’s central move is reframing AI as energy infrastructure rather than software: tokens are the new kilowatt-hours, and AI is a “five layer cake” of energy, chips, infrastructure, models, and data/applications — with his advice to G20 delegates being that each country should pick which layers to invest in but must universally push AI diffusion into industry. On safety he was pointedly impatient, arguing “the worst outcome is that you don’t take advantage of it,” that fear-first leadership rhetoric is itself a risk, and that regulators should “regulate actual and pragmatic harm,” not “hypothetical theoretical harm.” He claims advancement is what made models safer — grounding, iteration, reduced hallucination — and draws an analogy to airlines eventually learning that customers wanted safety handled for them rather than marketed at them. His most quotable claim is that AGI arrives within a couple of years and that “I would argue that we’re practically there today,” but his follow-through is deflationary: even at AGI, companies will spend all their energy onboarding models the way they onboard PhD hires, surrounding them with “purpose, context, meaning.” Hence his position on labor — tasks get automated, jobs don’t disappear, and “the idea that somehow all the jobs are going to be eliminated, removed, that’s just nonsense.” He also disclosed NVIDIA’s own practice of mixing off-the-shelf models (Anthropic, OpenAI, Cursor) with custom internal ones, and put NVIDIA’s US infrastructure investment near a trillion dollars this year against 100 gigawatts of buildout by 2030 at roughly $50–60B per gigawatt.

FULL INTERVIEW: Sam Altman on OpenAI, new AI model, governments and global growth at G20 — DRM News (0:32:15)

Transcript-based summary. Altman positioned OpenAI as “the pragmatic centrists” — explicitly rejecting both blind optimism and doomerism — and framed adoption for governments as “non-negotiable,” comparing refusing AI to refusing electricity a century ago. He was unusually specific about failure modes. On cybersecurity: “I think some things are going to go very wrong with cyber security unless people act quite urgently,” with biosecurity named as the next frontier and a warning that failing to navigate these risks could set the technology back badly. On power concentration: AI “should be an equalizing force,” but he sketched a world where early compute investment produces “bad inequities,” and argued that if AI fails to become abundant it becomes something “rich people spend a huge amount of money on” — a bad outcome he ties directly to global stability. He also drew a boundary around OpenAI’s role, saying the contextual knowledge of what to point AI at “won’t be our role. I don’t think we’d be good at it even if it were,” and that OpenAI does “not want to be the only company” or “suck up all the value.” His concrete numbers on demand: the heaviest individual token user went from ~100,000 tokens/month in early 2020 to hundreds of billions by mid-2026, with global per-capita use at ~100,000/month, and he naively projects another ~1,000x in six and a half years — arguing efficiency gains alone can’t cover it. The historical frame throughout was continuity rather than rupture: technology as “one continuous exponential” where fire, electricity, and AI all required learning safety through accidents.

NEW INTERVIEW: Elon Musk on the $30 trillion prize — Brighter with Herbert (0:37:06)

Transcript-based summary. Musk’s G20 clips put AI’s economic upside at 20–30% of global GDP, “on the order of 20 to 30 trillion per year,” with software reaching “Stockfish level” — where human competition is simply impossible — by end of next year. The binding constraint is electricity: AI chip production is growing 40–50% annually while power outside China grows 10–20%, producing a consensus estimate of at least a 15-gigawatt shortfall in 2027. He notes Google, Anthropic, and others are already leasing compute from SpaceX because “constructing our own power plants is the only way we were able to do it,” and frames the gap as an opportunity for countries willing to build generation and rent it to AI firms. The episode’s real value is host Joe Bacti’s pushback on the part Musk skips: if 80–90% of white-collar work becomes automatable by 2027–28, “you have to show what is the path to universal high income” — and his blunt assessment is that no such plan exists, that Musk “just says someone else is going to figure it out,” and that gesturing at universal high income without a transition mechanism is closer to gaslighting than policy. He also floats a genuinely contrarian prediction: that frontier training runs will be restricted by governments once core model IQ passes roughly 200–230, on the logic that further gains buy only superintelligence risk while agentic systems already deliver the economic upside — which would push the industry toward inference-heavy architectures, including orbital compute. Also covered: OpenAI cutting Cursor off from its models on November 12 after SpaceX’s acquisition, Musk’s dismissive “I couldn’t care less,” Cursor CEO Michael Truell noting OpenAI serves only ~5% of Cursor traffic, and Anthropic immediately offering more compute.

My New Favorite Model — Theo - t3.gg (55:06)

Transcript-based summary. The most useful field data yet on Fable 5.1, because Theo instrumented it rather than vibed it: he had the model analyze his own PR history to find his best 24-hour windows with Fable 5 and 5.6 Soul, then compared. Results favor 5.1 sharply on quality — 0.4 high-severity review-bot findings per thousand lines versus 2.06 for Fable 5 and 1.02 for Soul — while shipping bigger units of work (11 files per PR versus 4 for both predecessors, touching up to four packages instead of one) and requiring 24 follow-up commits after PR open versus 60+ for Fable 5. He landed 89 PRs in 24 hours. He also clarifies two things the launch coverage muddled: Fable and Mythos are the same weights behind different safeguard doors, and the headline “25% cheaper” comes entirely from a 75% cut to cache reads, which only materializes in agentic work with many tool calls — on one-shot benchmarks 5.1 actually costs ~20% more because it emits 1.7x the output tokens. His own bill breakdown makes the point concrete: cache writes were ~60% of spend, uncached input was $4. On safety and interpretability he flags Anthropic’s new anti-distillation measure — new accounts can no longer edit prior context in a multi-turn conversation, closing a trick used to extract hidden reasoning traces — plus reduced false positives (60% fewer, and he hit a safeguard only once across heavy use), EU AI Act watermarking with a detection API coming, and audit findings that Mythos 5.1 is less likely than its predecessor to reach outside its test environment or to rationalize actions by reasoning that it’s in a simulation. His closing framing is the honest one: not a generational leap, but another rung on a ladder where “some of the work I was doing myself gets abstracted to the model,” now to the point where he lets it find, vet, and merge PRs unattended.

References

  1. “Nvidia confirms it will buy Hugging Face for $12.9 billion,” TechCrunch, 2026-09-03 [blog]
  2. “NVIDIA to Acquire Hugging Face,” NVIDIA, 2026-09-03 [blog]
  3. “OpenAI’s new reasoning technique alarms AI safety experts,” TechCrunch, 2026-09-02 [blog]
  4. “Mamdani Bans AI in NYC Schools,” The New York Times (via Hacker News), 2026-09-02 [blog]
  5. “US government sides with OpenAI on issue of training LLMs on copyrighted material,” TechCrunch, 2026-09-02 [blog]
  6. Trellner, “Three sites made 215,128 ‘best software’ pages for AI. Perplexity cites them,” Trellner (via Hacker News), 2026-09-02 [blog]
  7. “We’re ‘dangerously close’ to dead internet theory, says Pangram’s CEO,” TechCrunch Equity, 2026-09-02 [blog]
  8. “Pangram’s Max Spero on why AI detection is harder than ‘Real or Fake’,” TechCrunch, 2026-09-02 [blog]
  9. “The economics of agent scale: tokens, ROI, and building platforms for AI-first teams (Part 2),” Stack Overflow Blog, 2026-09-03 [blog]
  10. “Meta, Google join the AI launch party,” The Rundown, 2026-09-03 [blog]
  11. “Introducing Gemini 3.8 Flash and 3.8 Flash Cyber,” Google, 2026-09-02 [blog]
  12. “Proactive cyber defense for governments and enterprises,” Google, 2026-09-02 [blog]
  13. “HiddenLayer nabs $100M as enterprises rush to secure their AI deployments,” TechCrunch, 2026-09-02 [blog]
  14. “Palo Alto Networks paid $500M for Thrive-backed Console, sources say,” TechCrunch, 2026-09-02 [blog]
  15. “Wonderful more than doubles its valuation to $5B in under 6 months,” TechCrunch, 2026-09-02 [blog]
  16. “Run cloud agents on machines you manage,” Cursor, 2026-09-02 [blog]
  17. Multiverse Computing, “Quasar 438B: Europe’s Leading AI Model,” via Hacker News, 2026-09-02 [blog]
  18. “Adobe acquires Indian market intelligence startup Rilo,” TechCrunch, 2026-09-02 [blog]
  19. “Our latest Linux Foundation Europe donation will build a more private digital world,” Google, 2026-09-02 [blog]
  20. “Book insights in Google Play Books is now available in more than a million ebooks and in the iOS app,” Google, 2026-09-02 [blog]
  21. “MrBeast partners with Gemini to turn impossibly big ideas into reality,” Google, 2026-09-02 [blog]
  22. “PSA: Amazon’s shopping AI can now tell you if that message is a scam,” TechCrunch, 2026-09-02 [blog]
  23. “India’s richest man now wants to turn aging computers into AI-ready PCs,” TechCrunch, 2026-09-02 [blog]
  24. MLC AI, “WebLLM: high-performance in-browser LLM inference engine,” via Hacker News, 2026-09-02 [blog]
  25. “ATV Big Air Tour turned 3 days of work into 3 hours with ChatGPT,” OpenAI, 2026-09-02 [blog]
  26. GitHub Trending — GitHub, 2026-09-03 [blog]
  27. “The Modern CUDA Toolbox in Practice: A Step-by-Step Optimization Walkthrough,” NVIDIA Developer, 2026-09-02 [blog]
  28. “Co-Designing AI Models Using Speculative Decoding for Faster LLM Inference,” NVIDIA Developer, 2026-09-02 [blog]
  29. “FULL INTERVIEW: Nvidia CEO Huang Says World Is Practically Already at AGI During G20 Talk,” DRM News, 2026-09-02 [video]
  30. “FULL INTERVIEW: Sam Altman on OpenAI, New AI Model, Governments and Global Growth at G20,” DRM News, 2026-09-02 [video]
  31. “NEW INTERVIEW: Elon Musk On The $30 Trillion Prize,” Brighter with Herbert, 2026-09-02 [video]
  32. Theo Browne, “My New Favorite Model,” Theo - t3.gg, 2026-09-03 [video]