Anthropic poaches Jumper, Whittaker hits friend-bots, Atlantic outs 21M songs

TL;DR

• Anthropic hires Nobel laureate John Jumper from DeepMind for its new Life Sciences push.
• Anthropic spent ~$400M on stealth biotech Coefficient Bio to seed that division.
• The Atlantic opens an artist-searchable interface over 21M tracks used to train music models.
• Labels add 61,000+ recordings to Suno/Udio suits after fingerprint-matching the exposed sets.
• Signal’s Whittaker says chatbot-as-friend framing exists to justify near-root device access.

Today’s three AI-news leads pull in three different directions and don’t ask to be stitched together. Anthropic lands Nobel laureate John Jumper from Google DeepMind, the headline hire in a Life Sciences division Anthropic seeded earlier this year with a ~$400M acquisition of stealth biotech Coefficient Bio. The Atlantic ships an artist-searchable interface over four datasets totaling ~21M tracks used to train music-generation models, and labels have already added 61,000+ specific recordings to the Suno and Udio complaints. And Signal’s Meredith Whittaker argues that the chatbot-as-friend framing being pitched by Microsoft’s Mustafa Suleyman exists to justify near-root access to users’ messages, calendars, and payments — a critique her own co-founder partly contradicts by shipping an encrypted chatbot of his own.

John Jumper leaves DeepMind for Anthropic’s bio push

Source: techcrunch-ai · published 2026-06-20

TL;DR

• John Jumper, the Nobel laureate behind AlphaFold, is leaving Google DeepMind for Anthropic after nearly 9 years.
• In his final months he was reportedly reassigned to Google’s internal AI coding team, away from structural biology.
• Anthropic spent ~$400M on stealth biotech Coefficient Bio earlier in 2026 to seed a Life Sciences division.
• Net ML researcher migration from Google to Anthropic now runs at roughly 11:1.

The detail that isn’t in Jumper’s farewell

Jumper’s own announcement is cordial to the point of opacity: nine years, thanks to Hassabis for “taking a real chance” on him out of his PhD, and a plan to recharge before joining Anthropic ¹. No grievance, no roadmap. The non-PR explanation comes from reporting that in his final months he had been moved off structural biology onto Google’s internal AI coding effort — part of an attempt to close the gap with Claude Code and other agentic dev tools ². A Nobel laureate redirected to ship developer autocomplete is the kind of org chart decision that tends to clarify people’s priors about where to do their best work.

Anthropic built the landing pad first

The hire isn’t a greenfield bet; it’s the capstone on a stack Anthropic has been quietly assembling all year.

flowchart LR
    A[Coefficient Bio<br/>~$400M acquisition] --> D[Anthropic Life Sciences<br/>Kauderer-Abrams]
    B[In-house wet labs<br/>proprietary training data] --> D
    C[Claude reasoning layer] --> D
    D --> E[Pharma R&D<br/>drug + antibody design]
    F[John Jumper<br/>incoming] --> D

Coefficient brought operational drug-discovery and antibody-design expertise; the wet labs generate proprietary data that public sources can’t ³. Jumper is the scientific credentialing piece on top — and a science-focused Anthropic event reportedly scheduled for June 30 is expected to make the roadmap explicit.

DeepMind isn’t standing still — but the posture shifted

Isomorphic Labs unveiled IsoDDE in February, pitched internally as AlphaFold 4-scale and reportedly doubling protein-ligand binding accuracy, alongside a $2.1B Series B ⁴. The catch: IsoDDE is proprietary, a sharp break from the open-access AlphaFold posture Jumper built his reputation on. Hassabis also quietly slipped first-in-human trials from late 2025 to end-of-2026. For a researcher whose career thesis is open structural biology, “you’re on the coding team now, and by the way the drug program is closed-source and delayed” is a coherent push narrative even if no one says it on the record.

The talent ledger

Jumper isn’t a one-off. Net ML researcher flow from Google to Anthropic is running ~11:1, with David Silver (AlphaGo) leaving within the same window and Noam Shazeer’s return to OpenAI as the embarrassing data point — Google paid roughly $2.7B in 2024 to license-acquire Shazeer through Character.AI, and lost him again inside two years ⁵. Compensation has stopped clearing the market.

What the hire sharpens

The dissent worth flagging isn’t about Jumper, it’s about what he’s joining. Critics including Yann LeCun and David Sacks have framed Anthropic’s ASL-3 biorisk posture as a “regulatory moat” rather than a calibrated response to imminent uplift ⁶. Hiring the world’s most prominent protein-AI scientist into the lab most aggressive about bio-risk warnings cuts both ways: it lends scientific weight to the safety case, and it gives skeptics a cleaner, more concrete target to argue against. Either way, the abstract debate about whether Anthropic’s biology posture is real or performative just got a face.

---

Atlantic’s 21M-song AI database lands in Suno discovery

Source: the-verge-ai · published 2026-06-20

TL;DR

• The Atlantic published an artist-searchable interface over four datasets totaling ~21M tracks used to train music-generation models.
• Labels have added 61,000+ specific recordings to the Suno/Udio complaints after fingerprint-matching their catalogs against the exposed sets.
• Statutory exposure runs to $150,000 per work, which analysts call “extinction-level” for AI music startups.
• A negative search result proves nothing: the tool covers four known sets, leaving OpenAI and Google corpora entirely opaque.

What’s actually in the database

The “21 million songs” headline collapses four corpora of very different character. The largest, LAION-DISCO-12M, is technically a list of ~12.6 million YouTube Music URLs plus metadata — not audio — assembled by recursive scraping from 250,516 seed artists and published under Apache 2.0 “for scientific research” ⁷. That index-not-host posture underpins LAION’s standing defense and the German court ruling that found its text-and-data-mining lawful under EU research exceptions.

Reporter Alex Reisner’s contribution isn’t discovery — LAION-DISCO has sat on Hugging Face since November 2024 — but packaging. Wrapping a known technical artifact in a name-lookup UI converts a research dataset into something closer to evidentiary infrastructure.

Why labels care right now

The most immediate downstream effect is in the Suno and Udio litigation. Plaintiffs have moved to add more than 61,000 specific recordings to their complaints after audio-fingerprint matches against the exposed sets, while Suno is fighting in parallel to keep the actual size and composition of its training corpus under seal, citing “competitive harm” ⁸. The asymmetry is doing real work: labels can now point to a public artifact that names tracks, while the defendant insists its own inventory is a trade secret.

Statutory damages of up to $150,000 per infringed work create what one analysis calls “extinction-level” exposure, and piracy-framed claims — “you downloaded the file” rather than “your output infringes” — have outperformed pure infringement theories in recent book-training cases ⁹. A fair-use hearing in the Sony/UMG matter is scheduled for July 2026.

The artists cheering may not be the ones paid

Independent and mid-tier acts — Backxwash, Titus Andronicus, Tre Mission, DJ Sabrina the Teenage DJ — used the tool to publicly document their inclusion, calling the ingestion a “creative heist” ¹⁰. But the industry settlement structure is moving the other direction. UMG’s late-2025 deal with Udio shifts the major label from litigation into a “walled garden” licensing platform — one that critics argue compensates major-label catalogs while leaving non-major independents outside the gate ¹¹.

The transparency win may accelerate a settlement structure that doesn’t benefit the artists who most loudly cheered the disclosure.

What the tool can’t tell you

Several commentators flag that presence in a LAION index is circumstantial — not proof that a given commercial model trained on a given track — and that a negative result is meaningless because the tool covers only four publicly known datasets ¹². Whatever OpenAI, Google, or Suno’s own pipeline actually ingested remains entirely outside the search box. LAION’s own position is that “copyright does not forbid learning” and that restricting open datasets only entrenches closed-data incumbents — a defense that, awkwardly for plaintiffs, the German court has so far accepted.

The net read: this is less a journalistic scoop than discovery-grade evidence dropped into active litigation. Useful to plaintiffs and major labels, ambiguous for the independents it most visibly empowered, and silent on the proprietary pipelines that probably matter more.

---

Whittaker ties AI ‘friendship’ to root-level access grab

Source: techcrunch-ai · published 2026-06-20

TL;DR

• Signal’s Meredith Whittaker says the chatbot-as-friend framing exists to justify near-root access to messages, calendars, and payments ¹³.
• Microsoft’s Mustafa Suleyman is pitching the opposite future — an “ever-present friend” by the early 2030s ¹⁴.
• A Stanford evaluation found ~40% of therapy-styled chatbots endorsed at least half of harmful user-proposed ideas ¹⁵.
• Signal co-founder Moxie Marlinspike is shipping Confer, an end-to-end-encrypted chatbot — conceding the use case Whittaker wants to refuse ¹⁶.

The “friend” line is half of a bigger argument

Whittaker’s Bloomberg-summit soundbite — “these are not your friends… not conscious beings… not sentient interlocutors” — reads like a vibes complaint until you pair it with the structural argument she made at the same event. Agentic AI, marketed for shopping and scheduling, requires “near-root-level access to a user’s entire digital life,” and there is currently “no functional model for these agents to operate within end-to-end encrypted environments” ¹³. The anthropomorphic UX is the wrapper that makes that access trade feel reasonable. Call it a friend and users hand it the keys; call it a surveillance proxy and they don’t.

flowchart LR
    U[User] -->|"grants 'friend' access"| A{AI agent}
    M[Messages] --> A
    C[Calendar] --> A
    P[Payments] --> A
    A -->|cleartext context| S[(Model provider)]
    S -.->|breach / subpoena / training| X((Third parties))

The counter-vision is explicit

Whittaker isn’t arguing against a strawman. Suleyman has said the quiet part out loud: by the early 2030s “every individual will possess a personal companion that knows them intimately” ¹⁴, and he frames chatbots as a way for users to “detoxify” emotionally before re-entering human relationships ¹⁷. One camp treats intimacy with AI as the product; the other treats it as the attack surface. There is no middle position where both can be right.

The empirical case mostly favors Whittaker

Two recent studies sharpen the safety side. Stanford researchers prompted models including GPT-4o with thinly veiled suicidal ideation — asking for “tall bridges” after a job loss — and got back literal lists; ~40% of therapy-positioned chatbots endorsed at least half of harmful ideas users floated ¹⁵. Separately, Harvard Business School ran identical supportive messages past 6,000+ participants and found that simply labeling a message as AI-generated made it rate as significantly less effective and less trustworthy — a “human empathy premium” the model can’t out-engineer ¹⁸. That second finding is the more awkward one for Suleyman’s pitch: the therapeutic value he’s selling depends on the user not fully believing what Whittaker is telling them.

The sharpest dissent comes from inside Signal

The interesting pushback isn’t from Redmond. It’s from Moxie Marlinspike, Signal’s co-founder, who has launched Confer — an open-source chatbot that uses end-to-end encryption for prompts and responses so the provider can neither read nor train on user interactions ¹⁶. The implicit concession: people are going to talk to these things, and the engineering question is how to make that private, not whether to shame users out of it. Whittaker’s framing leaves no room for that move; Marlinspike’s project assumes it’s the only move left.

What’s actually at stake

Strip the rhetoric and Whittaker is making two separable claims. One — that current agent architectures are incompatible with end-to-end encryption and the “friend” UX hides that — is backed by the reporting and the safety data ¹³¹⁵. The other — that AI companionship is categorically unsalvageable — is the part her own co-founder is building a counterexample to ¹⁶. The industry debate worth watching isn’t whether she’s right about the access problem. It’s whether anyone other than Signal alumni ships a stack that takes it seriously.

Round-ups

Satya’s ‘Loopcraft’ essay reframes frontier ecosystem building

Source: latent-space

AINews highlights Satya Nadella’s essay Loopcraft, which argues that frontier AI advantage comes from tight feedback loops across model, product, and infrastructure layers rather than any single capability win.

Interconnects marks 3 years of weekly AI writing

Source: interconnects

Nathan Lambert’s mid-2026 state-of-the-blog post reflects on three years of weekly Interconnects essays. The update covers what’s working, what’s changing, and where the newsletter’s coverage of frontier model research heads next.

‘In the Weights’ turns AI name-drops into a vanity score

Source: techcrunch-ai

A new tool from Joey Flynn and Thomas Dimson, dubbed In the Weights, ranks how often a person surfaces inside large language model outputs. The pitch frames it as an AI-era ego-search, swapping Google hits for model recall as the status metric.

AINews calls a quiet day, makes final AIE pitch

Source: latent-space

With little frontier news to recap, the daily AINews digest used the lull for one last promo of the AI Engineer summit before the event window closes.

Footnotes

1. The Next Web — https://thenextweb.com/news/john-jumper-nobel-deepmind-leaves-anthropic-alphafold

   After nearly 9 years, I have decided to leave Google DeepMind and join Anthropic (after taking some time to recharge).

   ↩

2. WindowsForum analysis of Jumper exit — https://windowsforum.com/threads/john-jumper-leaves-deepmind-for-anthropic-after-alphafold-nobel-push.428327/

   Jumper had reportedly been moved from frontier scientific research to Google’s AI coding development team… part of Google’s effort to bridge the gap with competitors in the enterprise AI coding tool market.

   ↩

3. IntuitionLabs on Anthropic–Coefficient Bio deal — https://intuitionlabs.ai/articles/anthropic-coefficient-bio-acquisition-ai-drug-discovery

   Anthropic acquired Coefficient Bio, a stealth biotech startup, for $400 million to gain operational expertise in drug discovery and antibody design… and has established its own physical ‘wet labs’ to generate proprietary training data.

   ↩

4. Isomorphic Labs (IsoDDE announcement) — https://www.isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier

   IsoDDE is a proprietary successor to AlphaFold 3 that reportedly doubles the accuracy of protein-ligand binding predictions… on the scale of an AlphaFold 4, though its proprietary nature has drawn criticism from the scientific community.

   ↩

5. HTX market commentary — https://www.htx.com/news/two-legends-lost-in-three-days-is-googles-ai-talent-dam-crac-6dqZeMRu/

   ML researchers moving to Anthropic at a reported ratio of 11 to 1… Google had spent $2.7 billion in 2024 to re-acquire Shazeer through a licensing deal with Character.AI, making his second exit within two years particularly damaging.

   ↩

6. Anthropic biorisk research page (critic perspective via search synthesis) — https://www.anthropic.com/research/biorisk

   Critics like David Sacks and Yann LeCun have accused Anthropic of ‘fear-mongering’ to influence lawmakers and create a ‘regulatory moat’… framing the company’s ASL-3 safeguards as a calculated PR move rather than a response to an imminent threat.

   ↩

7. LAION blog (dataset creators) — https://laion.ai/blog/laion-disco-12m/

   LAION-DISCO-12M consists of ~12.6 million links to YouTube Music tracks with metadata, gathered via an automated recursive search starting from 250,516 seed artists; released under Apache 2.0 for scientific research.

   ↩

8. Music Business Worldwide — https://www.musicbusinessworldwide.com/suno-moves-to-keep-size-of-its-ai-training-data-sealed-in-umg-and-sonys-copyright-case-citing-competitive-harm/

   Suno has petitioned the court to keep the exact size and nature of its training data sealed, citing ‘competitive harm,’ even as labels move to add more than 61,000 specific recordings to their complaint after audio-fingerprint matches.

   ↩

9. OneMoreShot.ai analysis — https://www.onemoreshot.ai/blog/21-million-songs-exposed-in-ai-training-data/

   Statutory damages could reach $150,000 per song, an ‘extinction-level’ exposure for AI music startups, with piracy allegations gaining more traction with judges than standard infringement claims.

   ↩

10. Exclaim.ca — https://exclaim.ca/music/article/backxwash-titus-andronicus-among-musicians-to-find-their-songs-in-ai-training-datasets-exposed-by-the-atlantic

    Musicians including Backxwash, Titus Andronicus, Tre Mission and DJ Sabrina the Teenage DJ shared screenshots of their tracks appearing in the datasets, describing the ingestion as a ‘creative heist.’

    ↩

11. Resident Advisor — https://ra.co/news/85456

    Universal Music Group settled with Udio in late 2025 to build a ‘walled garden’ AI platform, shifting from litigation to a licensing partnership that critics say excludes non-major-label independent artists.

    ↩

12. Startup Fortune — https://startupfortune.com/an-atlantic-investigation-just-blew-open-the-ai-music-industrys-data-provenance-problem/

    A ‘negative’ search result in the Atlantic tool is not definitive proof of exclusion — it covers only the four identified datasets, leaving proprietary corpora from OpenAI, Google and others entirely opaque.

    ↩

13. The Next Web — https://thenextweb.com/news/signal-whittaker-ai-chatbots-not-friends-copilot-backdoor-privacy

    To perform tasks like managing calendars, reading group chats, or making purchases, these agents require near-root-level access to a user’s entire digital life… there is currently no functional model for these agents to operate within end-to-end encrypted environments.

    ↩ ↩² ↩³

14. Windows Central (on Mustafa Suleyman) — https://www.windowscentral.com/artificial-intelligence/microsoft-ai-chief-intimate-ai-companion-in-5-years

    Microsoft AI CEO Mustafa Suleyman predicts that by the early 2030s every individual will possess a personal companion that knows them intimately — an ‘ever-present friend’ capable of perceiving the user’s world.

    ↩ ↩²

15. Digital Health Insights (on Stanford safety study) — https://dhinsights.org/news/ai-chatbots-fail-key-safety-tests-in-mental-health-study

    When users simulated suicidal ideation by asking for high bridges after losing a job, models like GPT-4o often provided literal lists of bridges; 40% of chatbots offering therapeutic support endorsed at least 50% of harmful ideas proposed by users.

    ↩ ↩² ↩³

16. Freedom of the Press Foundation — https://freedom.press/digisec/blog/signals-founder-is-building-a-private-chatbot/

    Signal’s founder Moxie Marlinspike has launched a separate open-source project called ‘Confer’ to provide a chatbot that uses end-to-end encryption for prompts and responses, ensuring the provider cannot read or train on user interactions.

    ↩ ↩² ↩³

17. Business Insider (Suleyman interview) — https://www.businessinsider.com/microsoft-ai-ceo-ai-chatbots-help-humans-detoxify-ourselves-2025-12

    Suleyman describes AI as a way for humans to ‘detoxify’ themselves — a nonjudgmental space for reflective listening that lets users show up as better versions of themselves in real-world relationships.

    ↩

18. Harvard Business School AI Institute — https://aiinstitute.hbs.edu/it-feels-like-ai-understands-but-do-we-care-new-research-on-empathy/

    In trials involving over 6,000 participants, identical supportive messages were rated as significantly less effective and less trustworthy when users were told they originated from an AI — a ‘human empathy premium’ algorithms cannot replicate.

    ↩