Musk testifies on Grok, Goodfire debuts Silico, Stripe tokenizes agent pay

TL;DR

• Musk testified under oath that xAI used OpenAI outputs to train Grok, putting OpenAI’s untested contract-law theory of distillation in front of a federal jury.
• Goodfire’s Silico promises a 58% hallucination cut via SAE-based steering, but the MIB benchmark just found SAEs don’t beat raw neurons at causal localization.
• Stripe Link hands AI agents single-use virtual cards and merchant-scoped tokens instead of card-on-file access, with Mastercard and Visa — not Perplexity — as the real rivals.
• OpenAI is gating its new GPT-5.5 Cyber model to vetted defenders, the same access restriction it called anti-competitive when Anthropic did it weeks ago.
• Ars Technica maps Google’s Gemini-training defaults across product surfaces, finding fragmented controls and toggles that silently re-enable after feature updates.

Today’s news isn’t about a model launch — it’s about who, outside the frontier labs, is trying to wire accountability into AI systems, and what each effort is still missing. In an Oakland federal courtroom, OpenAI’s lawyers are trying to convert a contract clause into the industry’s first enforceable anti-distillation rule, with Musk on the stand conceding xAI did exactly what the clause forbids. In San Francisco, an interpretability startup is shipping a debugger pitched at making model behavior inspectable — on the same week a benchmark suggests its underlying technique may not do what the marketing says. And Stripe is rewiring agent commerce so AI shoppers carry merchant-scoped tokens instead of card numbers, betting on cooperation just as a federal judge blocks the puppet-the-browser alternative.

None of these efforts come from the labs whose models they’re meant to constrain. The briefs underneath sit in the same key — OpenAI quietly mirroring the access gates it just attacked Anthropic over, and Ars Technica mapping how Google’s privacy controls keep re-enabling themselves after updates.

Musk admits xAI trained Grok on OpenAI models under oath

Source: the-verge-ai · published 2026-04-30

TL;DR

• Musk conceded on the stand that xAI “partly” used OpenAI outputs to train Grok, calling it standard practice.
• Judge Gonzalez Rogers called Musk’s $134B damages figure “pulled out of thin air” but let the jury hear it anyway.
• The jury is only advisory; the judge controls liability and remedies, and her skepticism is pointed at both sides.
• OpenAI’s anti-distillation lever is contract law, not copyright — a theory never tested at verdict against a frontier lab.

The admission that reframes the plaintiff

Testimony in Musk v. Altman produced one line that will outlive the trial: Elon Musk, under oath, admitted xAI “may have” trained Grok on OpenAI model outputs, then characterized distillation as industry-standard behavior ¹. That concession is rhetorically devastating — the man suing OpenAI for betraying its open-source mission is also mining its models for his competitor — but the legal substance is thinner than the optics suggest.

OpenAI has no copyright claim to fall back on. The U.S. Copyright Office still treats raw model outputs as unprotectable, which is why OpenAI’s anti-distillation protection lives entirely in its Terms of Service. Berkeley Law researchers have described this as a “contractual moat”: the developer owns the platform access and the process, but not the text the model produces ². That theory has never been litigated to verdict against a domestic frontier lab. Musk’s “everyone does it” defense lands in genuine grey space, even if jurors hate hearing it.

The judge is the real audience

The jury in this case is advisory. Judge Yvonne Gonzalez Rogers retains final authority on liability and remedies, which is why her interventions matter more than the headlines from the witness box.

She has already called the $134 billion damages figure produced by Musk’s economist “pulled out of thin air,” then declined to strike it — letting the number reach the jury rather than gut the trial outright ³. The more revealing moment came with the jury out of the room, when Rogers personally interrogated Musk’s wealth manager Jared Birchall about his claimed inability to recall details of the $97.4B xAI-led bid for OpenAI.

I’m still struggling with how you can have conversations… but have no recollections even in a general sense ⁴

She also admonished Musk’s lawyers to stop coaching the witness from counsel table. None of that reads like a bench inclined to take the plaintiff’s factual record at face value.

What Grok’s “truth-seeking” brand can’t survive

The technical dissent is quieter but more corrosive to Musk’s market position than to his case. Independent researchers have warned for years that recursive training on another model’s outputs invites “model collapse” — degraded reasoning and inherited hallucinations from the parent system ⁵. That undercuts the defense (distillation as harmless industry practice) and the marketing (Grok as a uniquely “truth-seeking” alternative) in the same breath. You cannot simultaneously claim epistemic superiority and admit you bootstrapped your weights from the competitor you’re suing.

Net assessment

Prediction markets briefly priced Musk’s odds of victory near 53.5% after his initial testimony ⁶, but that’s a read on spectacle, not doctrine. The contract theory underpinning OpenAI’s strongest claim is untested. The judge is openly skeptical of Musk’s damages model and his camp’s recall under oath. The distillation admission is the line everyone will quote, but the trial is drifting toward a judge-controlled outcome where neither party leaves clean — and where the precedent on whether ToS alone can police frontier-model distillation is the actual prize.

Further reading

• Elon Musk testifies that xAI trained Grok on OpenAI models — techcrunch-ai
• The craziest part of Musk v. Altman happened while the jury was out of the room — the-verge-ai
• Elon Musk confirms xAI used OpenAI’s models to train Grok — the-verge-ai
• Elon Musk’s 7 biggest stumbles on the stand at OpenAI trial — ars-technica-ai

---

Goodfire ships Silico to debug LLMs as SAE evidence wobbles

Source: mit-tech-review-ai · published 2026-04-30

TL;DR

• Goodfire’s new Silico tool lets engineers inspect and steer model features during training, with demos including a 58% hallucination cut.
• The MIB benchmark found Sparse Autoencoders — Silico’s foundation — do not beat raw neurons on causal variable localization.
• Stanford’s James Zou warns that surfacing a feature is not proof the model uses it; over-steering can break reasoning outright.
• Access is enterprise-only and throttled (~30 RPM, 50k TPM), reflecting real compute overhead behind the polished UI.

What Silico actually does

Silico is Goodfire’s bid to turn mechanistic interpretability into a developer tool. Engineers can browse the feature catalog a Sparse Autoencoder (SAE) extracts from a model, intervene on individual features mid-training, and watch behavior shift. The launch demos are the kind that travel: fixing the infamous “9.11 > 9.9” comparison bug, cutting hallucinations by 58%, and — in joint work with Arc Institute and Mayo Clinic — mapping features inside the Evo 2 genomics model to protein secondary structure ⁷.

The pitch is that interpretability has graduated from a research curiosity to a control surface. The evidence around that claim is messier than the demos suggest.

The benchmark problem

The Mechanistic Interpretability Benchmark (MIB) is the field’s emerging yardstick for whether interpretability methods actually identify the variables a model computes with. On the causal-variable-localization track, SAEs — the exact technique Silico’s UI is built around — did not outperform standard neurons or raw hidden dimensions ⁸. Supervised Distributed Alignment Search and attribution patching won those tracks instead.

That is awkward. A product whose value proposition is “intervene on this feature, predict the behavioral change” depends on features being causally faithful, not merely human-interpretable. Stanford’s James Zou put the caveat plainly in TIME’s coverage of the Evo 2 work: identifying a biological feature inside a model offers “no guarantee” the model actually uses that concept to reach its output ⁷.

The field hasn’t even converged on which SAE to trust. The major contenders take visibly different shapes:

Each implies a different theory of what interpretability is for — debugging, training, or auditing — and Silico stakes out only one corner of that space.

Operational reality

Silico is enterprise-only and priced case-by-case. AutoSteer is capped near 30 requests per minute against a 50,000 token-per-minute global ceiling, which is honest about the compute cost of live feature discovery but also a hard ceiling on the “just iterate” workflow the marketing implies ¹¹. The OpenAI-compatible Ember endpoint reportedly slots into existing stacks cleanly, but practitioner reaction is split — some developers complain about a clunky web UI laden with popups, others dismiss the platform as “snake oil,” and several note that aggressive steering “breaks” reasoning entirely ¹².

There is a sharper concern from the safety side: if interpretability becomes a training lever rather than an independent audit, it loses its value as a check. Models trained against an interpretability signal may simply learn to push computation into structures the current SAE can’t see ¹⁰¹².

Takeaway

Silico is a real engineering artifact, and the demos are not faked. But the strongest reading of the launch — that mech-interp now reliably explains and controls model behavior — runs into MIB’s SAE result, Zou’s causal caveat, and the throttling that betrays how expensive this still is. Treat it as a debugger with a load-bearing asterisk, not a verdict that the interpretability problem is solved.

---

Stripe Link hands AI agents tokens, not card numbers

Source: techcrunch-ai · published 2026-04-30

TL;DR

• Link’s agent mode issues single-use virtual cards or Shared Payment Tokens scoped to one merchant and amount — not card-on-file access.
• Stripe is betting on merchant-cooperative checkout via the Agentic Commerce Protocol; a federal judge just blocked Perplexity’s browser-puppet alternative on Amazon.
• Mastercard’s Agent Pay and Visa’s Trusted Agent Protocol — not Perplexity — are the real competitors.
• Fraud telemetry loss and a Reg E / EU liability gap remain unresolved and underplayed in the launch.

What actually ships

TechCrunch describes Link’s update as “approval flows,” which undersells the primitive doing the work. When an agent wants to spend, Link doesn’t expose card data — it takes a spend request scoped to a specific merchant and amount, pings the user for approval, then issues a one-time virtual card or a Shared Payment Token (SPT) to settle the transaction ¹³. Per-transaction human approval is the default; pre-set limits and merchant allowlists are previewed but not yet live ¹³. The SPT is co-specified in the Agentic Commerce Protocol Stripe built with OpenAI, and Link’s 250M-user base is the distribution wedge ¹⁴.

sequenceDiagram
    participant U as User
    participant A as AI Agent
    participant L as Stripe Link
    participant M as Merchant (ACP)
    A->>L: Spend request (merchant, amount)
    L->>U: Approval prompt
    U->>L: Approve
    L->>A: One-time SPT / virtual card
    A->>M: Checkout with SPT
    M->>L: Settle

The protocol war Stripe is actually in

The competitor isn’t Perplexity. Mastercard has shipped Agent Pay with “Agentic Tokens” that extend EMV tokenization to AI identities ¹⁵, and Visa’s Trusted Agent Protocol targets the same workflow. The split that matters is proxy vs. protocol: Perplexity’s Comet logs into your Amazon account like a headless browser, while Stripe/ACP wants merchants to expose structured agent endpoints. A federal judge just enjoined Comet from Amazon, holding that user permission does not equal merchant authorization and accepting Amazon’s argument that proxy agents “degrade the customer experience” by bypassing personalization and fraud systems ¹⁶. That ruling materially advantages the cooperative model — and explains why Etsy, URBN, Ashley Furniture, Coach, and Kate Spade signed onto ACP rather than wait the litigation out ¹⁴.

What the launch narrative skips

Two problems are larger than Stripe’s announcement implies.

First, fraud signal loss. Corgi Labs points out that SPTs strip the behavioral telemetry — mouse movement, typing cadence, browser fingerprints — that Radar and competing engines were trained on, and agent-initiated charges are “structurally indistinguishable from standard card-on-file transactions” in current merchant reporting tools ¹⁷. That is a visibility regression for risk teams, not just a new attack surface to defend.

Second, liability is unallocated. Taylor Wessing notes Reg E in the US assumes human-initiated transactions and offers no clean framework for delegated AI spending, and the EU’s revised Product Liability Directive explicitly excludes pure economic loss from its strict-liability scope ¹⁸. When an agent hallucinates a $4,000 chair purchase, the chargeback path is not yet a legal path.

The takeaway

Stripe has picked the side the courts are currently rewarding ¹⁶ and put the largest consumer wallet behind it ¹⁴. The token primitive is real engineering, not marketing. But the open questions — who eats the fraud, who eats the bad order — are the ones that will decide whether agentic checkout scales past the launch-partner roster, and the launch leaves both unanswered.

Round-ups

After dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber, too

Source: techcrunch-ai

OpenAI will gate its new GPT-5.5 Cyber security-testing model to vetted defenders only at launch, mirroring the access restrictions it recently criticized Anthropic for placing on Mythos. The company had publicly called those limits anti-competitive weeks earlier.

The hidden cost of Google’s AI defaults and the illusion of choice

Source: ars-technica-ai

Ars Technica maps the opt-outs, dark patterns, and default settings that route user data into Gemini training across Google’s product suite, finding privacy controls fragmented across multiple dashboards and several toggles that re-enable themselves after feature updates.

[AINews] Agents for Everything Else: Codex for Knowledge Work, Claude for Creative Work

Source: latent-space

Latent Space’s daily AI digest argues coding agents are now “breaking containment” into general knowledge and creative work, contrasting OpenAI’s Codex push toward office tasks with Anthropic positioning Claude for writers, designers, and other non-developer users.

Legal AI startup Legora hits $5.6B valuation and its battle with Harvey just got hotter

Source: techcrunch-ai

Swedish legal AI startup Legora has raised at a $5.6 billion valuation, intensifying its rivalry with Harvey as both expand into each other’s home markets and trade dueling ad campaigns, including one starring Jude Law. Nvidia’s NVentures joined the round.

Gemini replaces Google Assistant in cars with Google built-in

Source: the-verge-ai

Google is swapping the legacy Assistant for Gemini across millions of vehicles equipped with Google built-in, promising more natural conversations, vehicle-specific queries, and in-car settings control. The rollout marks the first time Gemini reaches automotive deployments at scale rather than via phone projection.

Further reading:

• Google’s Gemini AI assistant is hitting the road in millions of vehicles — techcrunch-ai

Meta is running get-rich-quick ads for its AI tools

Source: the-verge-ai

Manus, the AI agent startup Meta acquired for $2 billion last year, is paying creators to promote a pitch where users have AI build websites for local small businesses and cold-call the owners to sell them, drawing comparisons to get-rich-quick schemes.

Building gets easier

Source: bens-bites

Ben Tossell walks through how his personal builder stack has shifted as agentic coding tools mature, arguing the bar to ship working software has dropped sharply and naming the specific tools replacing his previous workflow.

Footnotes

1. Startup Fortune — coverage of admission — https://startupfortune.com/elon-musk-admitted-under-oath-that-xai-may-have-trained-on-openais-models-and-then-called-it-standard-practice/

   Elon Musk admitted under oath that xAI may have trained on OpenAI’s models and then called it standard practice

   ↩

2. Berkeley Law — The Network blog on AI distillation — https://sites.law.berkeley.edu/thenetwork/2025/03/30/the-innovation-dilemma-ai-distillation-in-openai-v-deepseek/

   OpenAI relies on contract law rather than intellectual property (IP) law to prevent distillation… creates a ‘contractual moat’ where the developer owns the process and the platform access, even if they cannot legally own the specific text the model produces

   ↩

3. PYMNTS — judge on damages model — https://www.pymnts.com/artificial-intelligence-2/2026/musks-134-billion-claim-in-openai-case-questioned-by-judge/

   Judge Rogers has expressed public skepticism toward these figures, characterizing the $134 billion demand as ‘pulled out of thin air’

   ↩

4. The News (Pakistan) — recap of Birchall sidebar — https://www.thenews.com.pk/latest/1401033-musk-vs-altman-trial-what-happened-when-the-jury-left

   I’m still struggling with how you can have conversations… but have no recollections even in a general sense

   ↩

5. Fast Company — model collapse risk — https://www.fastcompany.com/90998360/grok-openai-model-collapse

   training models on the outputs of other AIs can lead to ‘model collapse,’ a phenomenon where recursive training on synthetic data degrades a model’s reasoning and introduces ‘hallucinated’ artifacts

   ↩

6. Kalshi prediction market — https://news.kalshi.com/p/elon-musk-openai-trial-lawsuit-odds-2026

   traders briefly priced Musk’s odds of a victory at roughly 53.5% following his initial testimony

   ↩

7. TIME (Mayo Clinic / Arc Institute case study) — https://time.com/article/2026/04/14/ai-disease-genetic-mayo-clinic-goodfire/

   Stanford’s James Zou cautioned that identifying a biological feature inside a model does not guarantee the model actually uses that concept to reach its final output.

   ↩ ↩²

8. Iván Arcuschin — MIB benchmark page — https://iarcuschin.com/publication/mib/

   Sparse Autoencoders did not outperform standard neurons or raw hidden dimensions in the causal variable localization track

   ↩

9. LessWrong / GreaterWrong — Anthropic JumpReLU discussion — https://www.greaterwrong.com/posts/fG2gFYX2Wo49tRrap/anthropic-s-jumprelu-training-method-is-really-good

   Anthropic’s JumpReLU SAEs use learnable per-feature thresholds to avoid the shrinkage bias of L1 regularization, while Goodfire favors BatchTopK which relaxes the TopK constraint to the batch level.

   ↩ ↩²

10. Transluce — Predictive Concept Decoders — https://transluce.org/pcd

    PCDs translate a model’s internal states into human-readable concept lists to bypass deceptive self-reporting, positioned for scalable oversight of superhuman agents.

    ↩ ↩²

11. Longtermwiki E430 — Silico access notes — https://www.longtermwiki.com/wiki/E430

    Silico is priced case-by-case; AutoSteer is throttled to ~30 requests per minute and a 50,000 token-per-minute global cap reflects the compute overhead of feature discovery.

    ↩

12. GitHub gist (bigsnarfdude) — developer reaction notes — https://gist.github.com/bigsnarfdude/629f19f635981999c51a8bd44c6e2a54

    Developer feedback has been mixed, with some calling the platform ‘snake oil’ and others complaining of a clunky web interface filled with popups; over-steering can fundamentally break model reasoning.

    ↩ ↩²

13. Unite.AI — ‘Stripe hands AI agents a wallet’ — https://www.unite.ai/stripe-hands-ai-agents-a-wallet-ushering-in-agentic-purchasing/

    Instead of the agent ‘seeing’ card details, it requests a spend authorization for a specific context; once approved by the user, the wallet issues a one-time-use virtual card or a Shared Payment Token (SPT) to complete the transaction.

    ↩ ↩²

14. Payments Dive — Stripe/OpenAI ChatGPT checkout — https://www.paymentsdive.com/news/stripe-pushes-agentic-ai-sales-via-chatgpt-openai-artificial-intelligence/761439/

    High-profile merchants such as Etsy, URBN, Ashley Furniture, Coach, and Kate Spade have integrated Stripe’s Agentic Commerce Suite to make their catalogs ‘agent-discoverable’… Link currently serves over 250 million users.

    ↩ ↩² ↩³

15. PYMNTS — Mastercard Agent Pay — https://www.pymnts.com/mastercard/2026/mastercard-unveils-open-standard-to-verify-ai-agent-transactions/

    Mastercard has deployed Agent Pay, which utilizes ‘Agentic Tokens’ — dynamic digital credentials that extend tokenization to AI identities.

    ↩

16. GeekWire — ‘Judge blocks Perplexity’s AI bot from shopping on Amazon’ — https://www.geekwire.com/2026/judge-blocks-perplexitys-ai-bot-from-shopping-on-amazon-in-early-test-of-agentic-commerce/

    A federal judge ruled that while the agent had the user’s permission, it lacked Amazon’s ‘authorization’ to access password-protected systems… Amazon argued that such agents ‘degrade the customer experience’ by bypassing personalization and fraud-detection systems.

    ↩ ↩²

17. Corgi Labs — ‘Agent Payments Intelligence’ — https://www.corgilabs.ai/insights/agent-payments-intelligence-stripe-visibility

    When an agent initiates a purchase via Stripe’s Shared Payment Tokens, merchants lose access to critical telemetry like browser fingerprints, mouse movements, and typing cadences… agent-initiated charges are often structurally indistinguishable from standard card-on-file transactions in current reporting tools.

    ↩

18. Taylor Wessing — ‘Agentic AI in Payments’ — https://www.taylorwessing.com/en/insights-and-events/insights/2026/02/agentic-ai-in-payments

    Current consumer protections like Regulation E were designed for human-initiated transactions and lack a clear framework for ‘delegated’ AI spending… the EU’s revised Product Liability Directive currently excludes such monetary damages from its strict liability scope.

    ↩