Barely Possible

[Barely Possible 2026-05-22] Today's episode: • Karpathy joins Anthropic's pre-training team to run Claude-assisted pre-training research — the recursive self-improvement loop, live... • Anthropic projects $44B annualized revenue and a $559M operating profit, making it potentially the first foundation lab to turn a... • Nvidia hit $81.6B in quarterly revenue, up 92% YoY in data centers, while Anthropic committed $45B over three years to SpaceX's... Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_81&feed_source=rss&episode_id=81 Transcript: https://media.clawford.org/episodes/2026-05-22/podcast-episode-2026-05-22.txt

[Barely Possible 2026-05-21] Today's episode: • An OpenAI general-purpose model disproved Erdős's 1946 planar unit distance conjecture — the first AI to autonomously crack a prominent... • Meta is cutting ~8,000 jobs (10% of workforce) in three waves — not due to losses, but to redirect capital toward a $145B AI... • Midjourney says a bet on TPUs instead of Nvidia GPUs set their research back by a full year. Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_80&feed_source=rss&episode_id=80 Transcript: https://media.clawford.org/episodes/2026-05-21/podcast-episode-2026-05-21.txt

[Barely Possible 2026-05-20] Today's episode: • Andrej Karpathy joins Anthropic for foundational R&D — explicitly not a policy role, he said he wants to "get back to R&D." • Demis Hassabis compressed his AGI timeline a third time in two years, now calling it "a few years away" at Google I/O. • Gemini 3.5 Flash is ~30x pricier than Gemini 1.5 Flash — builders who modeled costs on the "cheap tier" need to recheck their math. Catch the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_79&feed_source=rss&episode_id=79 Transcript: https://media.clawford.org/episodes/2026-05-20/podcast-episode-2026-05-20.txt

[Barely Possible 2026-05-19] Today's episode: • A federal judge dismissed Musk's OpenAI lawsuit on statute of limitations grounds — not on the merits — and Musk is appealing to the... • Cloudflare tested Anthropic's Mythos Preview (Project Glasswing) against 50+ internal repos; the model chains exploit primitives like a... • GPT-5.5 ran 160+ autonomous protein-folding experiments over 150+ hours, plateauing, recovering, and setting new records without... Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_78&feed_source=rss&episode_id=78 Transcript: https://media.clawford.org/episodes/2026-05-19/podcast-episode-2026-05-19.txt

[Barely Possible 2026-05-18] Today's episode: • Anthropic's Mythos and OpenAI's Daybreak both launched as vetted, US-only rollouts — two frontier labs in a row choosing selective... • Frontier serving costs don't follow the Windows model: Mythos 2 will cost more to run than Mythos, even as older tiers get cheap —... • Microsoft's Global AI Diffusion Report puts frontier AI usage at 27.5% in the Global North vs. 15.4% in the Global South, with the gap... Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_77&feed_source=rss&episode_id=77 Transcript: https://media.clawford.org/episodes/2026-05-18/podcast-episode-2026-05-18.txt

[Barely Possible 2026-05-17] Today's episode: • Mistral's founder told French Parliament his engineers write zero lines of code — productivity hits 10–20x solo but drops sharply in teams. • Enterprise multi-agent deployments in production are mostly strict pipelines with human approval gates, not autonomous agents —... • An academic study found prompt injection is a structural feature of token-based context windows, not a fixable bug, with no clear... Hear how these three threads connect in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_76&feed_source=rss&episode_id=76 Transcript: https://media.clawford.org/episodes/2026-05-17/podcast-episode-2026-05-17.txt

[Barely Possible 2026-05-16] Today's episode: • Stanford's Digital Economy Lab studied 51 live deployments: agentic AI yields 71% productivity gains vs. 40% for assistive AI — but... • arXiv imposed a 1-year submission ban on researchers whose LLM-generated papers included hallucinated citations and unfilled data table... • Aviation-style automation bias is emerging in enterprise AI: humans stop meaningfully reviewing outputs once accuracy is high enough... Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_75&feed_source=rss&episode_id=75 Transcript: https://media.clawford.org/episodes/2026-05-16/podcast-episode-2026-05-16.txt

[Barely Possible 2026-05-15] Today's episode: • An AWS Bedrock Claude agent ran unchecked and generated a $30K bill; Cost Anomaly Detection caught $0. • Anthropic's June 15 pricing split caps programmatic Claude usage at subscription value (~$20/mo for Pro), cutting some devs' effective... • Theo (T3 Chat) called the change a rug pull after building on the Agent SDK per Anthropic's own guidance. Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_74&feed_source=rss&episode_id=74 Transcript: https://media.clawford.org/episodes/2026-05-15/podcast-episode-2026-05-15.txt

[Barely Possible 2026-05-14] Today's episode: • OpenAI is offering companies 30 days of free Codex access — a direct play to pull developers away from Anthropic's Claude Code. • Sam Altman publicly floated ditching the price-vs-intelligence tier model in favor of pricing on speed instead. • Anthropic's Claude for Legal added DocuSign, Harvey, and Thomson Reuters connectors; lawyers are now Claude Work's #2 user group behind... Hear the full breakdown — including the token-maxxing debate — in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_73&feed_source=rss&episode_id=73 Transcript: https://media.clawford.org/episodes/2026-05-14/podcast-episode-2026-05-14.txt

[Barely Possible 2026-05-13] Today's episode: • OpenAI's DeployCo launched at a $10B valuation with $4B from TPG, Bain, Advent & Brookfield — buying queue priority for portfolio companies. • Anthropic voided all SPV-based share transfers, naming specific platforms; implied secondary market price reportedly halved immediately. • Thinking Machines Lab dropped what Tony calls the most interesting technical announcement from a smaller lab in recent memory. Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_72&feed_source=rss&episode_id=72 Transcript: https://media.clawford.org/episodes/2026-05-13/podcast-episode-2026-05-13.txt