TL;DR
7 takeawaysAltman and Amodei briefed the UN Security Council on AI risks and called for international coordination; Trump the day before said the U.S. will accelerate AI, not rein it in.
Sources: CNBCThe GuardianFrance 24
Anthropic: ~950 Claude agents over 21 hours / 210M tokens found ART (array-associated reverse transcriptases)—a phage RT + partner gene + CRISPR-like repeat array expressed as short RNAs; function still unknown; Feng Zhang called it intriguing.
Sources: AnthropicThe VergeReuters
Google DeepMind: Private AI Compute gains persistent cross-device cloud memory in hardware enclaves; decryption keys stay on the user’s devices; tamper-proof software attestation + independent audit published.
Sources: DeepMind
Enveda: $311M Series E at $2B (Catalio-led; Iconiq et al.) to push AI-discovered, nature-derived drugs deeper into clinical trials, including skin and post-GLP-1 weight-maintenance candidates.
Sources: TechCrunch
Basecamp Research: $140M Series C led by S32 with NVIDIA and Anthropic’s Anthology Fund to train next EDEN models and advance in vivo cell-therapy pipeline.
Sources: PR Newswire
Outerlimit: $16M pre-seed for a decentralized discover/observe/enforce authorization layer that blocks off-policy agent tool actions regardless of model alignment.
Sources: SecurityWeek
Midcentury: $15M seed for egocentric robotics data (2M+ hours) plus Matrix cloud simulation for physical-AI policy eval.
Sources: tao.mediamidcentury.xyz
The Details
Story 01 / 06
Altman and Amodei take AI safety to the UN Security Council
- What happened
- On September 23, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei addressed the United Nations Security Council in separate briefings on AI safety. Altman argued that rival countries have historically cooperated around powerful new technologies and that AI must be one of those moments; he also said OpenAI has “unilaterally slowed down in the past” and will do so again, and called for international standards plus fast incident reporting. Amodei reiterated his multi-step pacing plan, including coordination across democratic and authoritarian governments, saying no leader, company, or nation can manage the risk alone. Their remarks came one day after President Donald Trump told the UN General Assembly the U.S. rejects a “globalist scheme” to control AI and will encourage progress rather than rein it in. France 24 quoted Altman urging “extreme care” as systems grow more autonomous, and Amodei pledging Anthropic will “slow down as much as necessary” for safe successive releases. Both companies are widely expected to pursue blockbuster IPOs even as they talk about pacing.
- Why it matters
- The public split is now institutional: lab CEOs are asking the Security Council for global coordination while the U.S. president frames international AI controls as a growth threat. Buyers and regulators will watch whether pacing talk constrains shipping—or only the messaging.
- Enterprise impact
- Expect more pressure for incident disclosure, third-party evals, and kill-switch / containment language in contracts—even if U.S. federal policy stays accelerationist. Map which vendors have actually paused training after containment failures.
- Startup impact
- Safety theater vs real process will be diligence questions. Document your eval harness, misuse refusals, and what you would pause for—investors and enterprise buyers will ask.
- What to focus on
- Read CNBC and France 24 for the UN quotes; separate “we will slow when needed” from any freeze on mid-tier model releases (yesterday’s Opus 5.5 / GPT-6 Sol still landed).
Discuss this story
Comment with your GitHub account. One thread per story.
Story 02 / 06
Claude finds ART—a CRISPR-like enzyme system in phages
- What happened
- Anthropic announced the first result from its new Bay Area life-sciences lab and Claude-powered research group: array-associated reverse transcriptases (ART), a previously uncharacterized bacteriophage system. Roughly 950 Claude agents spent about 21 hours and 210 million tokens surveying reverse transcriptases across a massive DNA database; one agent reading raw sequence next to an odd RT spotted a tandem repeat array “by eye,” then measured spacing, compared known RT systems, and filed a human-readable report. ART has three parts—the reverse transcriptase (previously noted in isolation), a partner gene, and a long array of evenly spaced non-coding DNA repeats that resemble CRISPR arrays and are expressed as distinct short RNAs. Anthropic does not yet know ART’s function; lab work is BSL-1/BSL-2 and done by human scientists. CRISPR pioneer Feng Zhang called the RNA-repeat/RT association “genuinely intriguing.” A preprint accompanies the blog; The Verge notes the early share is also a recruiting and capability signal ahead of IPO.
- Why it matters
- This is a concrete claim that autonomous LLM agents can initiate biological discovery—not just summarize papers—by noticing anomalies humans missed in sequence data. It also shows Anthropic pairing Claude with a real wet lab, not only software evals.
- Enterprise impact
- If you buy scientific AI, demand provenance: agent transcripts, novelty checks, and human lab confirmation—not press-friendly CRISPR analogies. Function of ART is still unknown.
- Startup impact
- Agent fleets that produce ranked hypothesis reports (then humans filter) are becoming a product pattern. The bottleneck shifts to scientific taste and wet-lab throughput.
- What to focus on
- Primary: Anthropic’s post and preprint. Treat CRISPR comparisons as structural analogy, not proven gene-editing utility.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 03 / 06
Google adds private persistent memory to Private AI Compute
- What happened
- Google DeepMind (with Platforms & Devices, Core, and Cloud) published a technical update to Private AI Compute: previously stateless hardware-isolated cloud enclaves will gain a persistent, cross-device memory layer. Personal context is stored encrypted in the cloud; cryptographic keys that unlock it remain exclusively on the user’s devices, so Google says even it cannot read the vault. When a model needs memory, an authenticated end-to-end encrypted channel opens a secure enclave that temporarily decrypts data in isolated memory, updates context, and re-encrypts. Google is also publishing a tamper-proof public record of server software for device attestation before any personal data is sent, plus results from an independent cybersecurity audit and an updated technical brief.
- Why it matters
- Long-horizon personal assistants need memory across phone, laptop, and glasses—but enterprises and regulators insist on on-device privacy norms. Device-held keys + attested enclaves are Google’s bid to square that circle.
- Enterprise impact
- Watch for B2B analogues: customer-controlled keys, enclave attestation, and auditability will become RFP line items for any “remembering” agent product.
- Startup impact
- Competing assistants that keep sticky memory in plaintext cloud stores will look weak. Design for sealed storage and client-side key custody early.
- What to focus on
- Read DeepMind’s Private AI Compute post and technical brief; this is architecture, not a consumer feature flip today.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 04 / 06
Enveda raises $311M at $2B for AI-discovered natural drugs
- What happened
- Enveda, which uses AI to find medicines in plants and microbes rather than synthesizing molecules from scratch, raised a $311 million Series E at a $2 billion valuation—double its valuation from 12 months earlier. Catalio Capital Management led; Iconiq and others participated. Founded in 2019 by Viswa Colluru (early Recursion hire), Enveda is among the AI-biotech cohort with candidates already in human trials, including programs for severe skin conditions and maintaining weight loss after stopping GLP-1 drugs. TechCrunch notes AI has still not produced an FDA-approved drug, so clinical progress is the proof point investors are buying.
- Why it matters
- Capital is concentrating on AI-biotech teams that clear the “in humans” bar, not just in-silico hit rates. Nature-derived chemical space is a differentiated data bet versus pure de novo design.
- Enterprise impact
- Pharma BD: watch which AI-native pipelines actually enroll and read out—not just raise. Post-GLP-1 maintenance is a high-visibility indication.
- Startup impact
- Clinic-ready assets + proprietary natural-product maps still raise at premium multiples. Discovery-only decks without IND paths will struggle next to Enveda.
- What to focus on
- TechCrunch Sep 23 write-up is the public primary for round size and valuation.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 05 / 06
Basecamp Research raises $140M to push EDEN therapeutics
- What happened
- London/Boston-based Basecamp Research announced a $140 million oversubscribed Series C led by S32 to train next-generation EDEN biological foundation models and advance AI-designed therapeutics toward the clinic, starting with in vivo cell therapy. Participants include NVIDIA, Anthropic’s Anthology Fund, Catalio, NATO Innovation Fund, Rockefeller Foundation, True Ventures, and others; Andy Conrad (S32; former Verily CEO) joins the board. Basecamp trains EDEN on its Trillion Gene Atlas—proprietary genomic data from access-and-benefit partnerships in 30+ countries—and pairs long DNA design with large serine recombinases for genomic integration. The company says it has strong preclinical results across modalities and appointed ex-Biogen Richard Pearce as CBO to expand pharma partnerships. It has also collaborated with Anthropic on Claude Science.
- Why it matters
- Another large check into AI-designed medicine, this time with chip (NVIDIA) and frontier-lab (Anthropic) capital on the same cap table as biotech specialists—signaling that biological foundation models are a strategic bet, not a niche academic toy.
- Enterprise impact
- If you partner on AI therapeutics, diligence the data rights story (ABS partnerships) and whether EDEN outputs are IND-trackable, not just pretty sequences.
- Startup impact
- In vivo cell therapy + proprietary atlases is the narrative. Expect more Foundation-model-for-biology raises with sovereign and defense-adjacent LPs.
- What to focus on
- Company PR Newswire release lists full investor set and pipeline focus.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 06 / 06
Outerlimit raises $16M to stop rogue agent actions
- What happened
- New York’s Outerlimit emerged from stealth with $16 million in pre-seed funding from AlbionVC, Evolution Equity Partners, Crane Venture Partners, and angels. Founders Tony Pepper, Neil Larkins, and Peter Vincent are building a decentralized authorization layer around discover, observe, and enforce: find agents in the environment, watch behavior, and block disallowed tool actions at execution time. The pitch is that alignment inside the model is insufficient—agents have no fear of consequences—so policy must bind identity, authorization, and action in one step, including scoped token use. SecurityWeek frames it as shifting from deterring humans to preventing harm from probabilistic agents operating at machine speed.
- Why it matters
- As enterprises grant agents credentials and tools, “prove the action is in policy before it executes” becomes the security product category—orthogonal to prompt filters.
- Enterprise impact
- Add agent discovery + runtime authorization to 2027 security roadmaps alongside IdP and secrets managers. Ask vendors for proofs that off-policy tool calls cannot complete.
- Startup impact
- If you ship agents with broad tool access, expect buyers to require an external policy layer. Outerlimit-style guardrails may become table stakes.
- What to focus on
- SecurityWeek Sep 23 feature is the clearest public description of the product thesis.
Discuss this story
Comment with your GitHub account. One thread per story.
Worth a Glance
Midcentury: $15M seed for physical-AI data + Matrix sim
Midcentury launched with a $15M seed for what it calls the largest egocentric robotics dataset (2M+ hours, 50+ environments, 20k tasks) and Matrix, a massively parallel cloud simulation loop to eval and improve robot policies before hardware.
tao.media(opens in a new tab)F13: $5M pre-seed for vector-graphics foundation models
F13 emerged from stealth with $5M (Credo, Point Nine) for controllable vector-graphics models aimed at professional design specs; early access via API and web app.
Tech.eu(opens in a new tab)
One Thing Not to Miss Today
Safety on the world stage—and AI that finds biology
Wednesday’s map: Altman and Amodei asked the UN Security Council for global AI coordination the day after Trump rejected international controls; Anthropic showed Claude initiating a real sequence-level biology find (ART) inside its wet lab; Google sketched how personal AI memory can live in the cloud with device-held keys; and biotech AI raised hard ($311M Enveda, $140M Basecamp) while agent-security (Outerlimit) and physical-AI infra (Midcentury) filled the builder stack. Watch whether UN rhetoric changes release cadence—or only the press cycle.