This morning’s first pass led with wet-lab protein binders: Claude orchestrated a 15-target campaign and 14 of them actually bound. That story still stands — recap at the bottom. The rest of the 24-hour tape is product ops and silicon capital: OpenAI’s teen surface, Google’s Marvell warrant, and the first MLPerf Client that scores agents on a PC.

Lead: ChatGPT for Teens is a default, not an app

OpenAI’s post (Aug 18, still rolling this week) is not a separate chatbot. If the system estimates someone is under 18, or they state they are 13–17, they are automatically dropped into ChatGPT for Teens. Help Center: availability began Aug 18 on Free and paid personal plans; Australia waits until September 8.

TechCrunch is the honest framing: teens have been on ChatGPT since 2022, lawsuits followed, and the product is arriving years after the user base. The operator question is not “is this nice.” It is “what actually changes when an account flips.”

The Under-18 model spec is the other half: no romantic language, no implied feelings, no emotional-dependence coaching. Break reminders, sensitive-image upload cautions, teen-specific onboarding. OpenAI also announced a CodeAI partnership for the pedagogy around the tool, plus the existing ChatGPT for Teachers path for classrooms.

Bold insight: this is the same subtractive product week as Microsoft pruning consumer Copilot. Big labs are no longer shipping every demo. They are shipping gated defaults. Age prediction plus auto-enroll is the actual mechanism — not a “Teen” button a 16-year-old can ignore.

TechCrunch’s caveat is the one to keep: teens are good at walking around parental controls. Until someone red-teams Study Mode the way people red-team jailbreaks, treat the homework-reminder claim as a product promise, not a measured cheat-reduction number.

If you are the adult in the loop this week — linking accounts, setting Quiet Hours, sitting next to a kid who is supposed to be doing algebra — the boring kit still wins: noise-cancelling headphones for the study block, a USB-C dock so the laptop is not dying mid-quiz, and a second monitor if you are reviewing the parental-control page while they work.

Secondary: Google just bought an option on Marvell’s TPU-attach stack

CNBC on Aug 19: Marvell jumped about 10% after a securities filing that lets Google buy up to 58,970,907 shares at 206.58 apiece — about 12.2 billion of equity, tied to purchase targets through fiscal 2033. Bloomberg adds the vest schedule: ~1.4 million shares vest in equal quarterly installments in year one; the rest vest on discretionary purchases, one tranche per 500 million of revenue from the co-developed products, from Marvell’s fiscal 2027 Q3 through fiscal 2033.

Marvell’s own language is the operator line: the expanded deal includes products that “attach to the TPU ecosystem” — inference accelerators, storage, network interface controllers. Google has spent a decade on custom silicon with Broadcom (that deal expanded again in April). Broadcom fell about 5% the same session. This is not “Google dumped Broadcom.” It is Google buying a second attach vendor and paying for it with a warrant that only vests if the purchases actually happen.

Same-day capital rhyme, already noted as a one-liner this morning and now with a primary: Nvidia is in talks on Mercor at a 20 billion valuation, double the October 10 billion Series C, with General Catalyst discussing the lead. Mercor is a data-labeling supplier Nvidia already buys from. Chipmaker taking equity in its own input. Warrant on the attach silicon, round on the labels. The circular-financing debate does not need another thinkpiece — just put the two term sheets on the same desk.

Tertiary: MLPerf Client v2.0 finally scores the agent, not the token

MLCommons shipped MLPerf Client v2.0 on Aug 18. The laptop benchmark used to be summarization, content, and code analysis. v2.0 adds the workloads people actually buy NPU laptops for:

AMD, Intel, Microsoft, NVIDIA, Qualcomm, and the usual OEMs are on the working group. Downloads and source: github.com/mlcommons/mlperf_client/releases.

Bold insight: a SWE-agent score that separates “the model thinking” from “the tools running” is the client-side cousin of yesterday’s AgentRadio story. If your NPU is fast and your tool loop is slow, the laptop loses on the part you can actually fix — sandbox, filesystem, retrieval — not the SKU.

Related, and stealable this week without buying a new box: IBM Research’s ALTK-Evolve memory-dose study (Aug 18). Eight models on AppWorld. Memory is a dose, not a feature flag.

Dumping every lesson into a small model costs ~50% more tokens and scores worse than a tight core plus per-task retrieval. If you are still stuffing the whole memory file into every ReAct step, that is the bug.

A Raspberry Pi 5 plus an NVMe enclosure is still the cheapest place to keep a small always-on agent warm while you run Client v2.0 on the actual laptop.

Recap: the tubes from this morning

Unchanged, still true, not the lead: Anthropic’s campaign hit binders on 14 of 15 targets. Adaptyv ran 1,320 designs; 354 bound (26.8%). MBP still produced zero. Protein design stays blocked in Fable 5. Liquid shipped QAD Q4_0 GGUFs for LFM2.5 230M–2.6B (~97% of BF16 recovered). Cerebras CS-4 is still a vendor 30× until independent serving configs land.

Takeaway

1. Today: if anyone in the house is 13–17 on ChatGPT, read the Help Center page and decide Study Hours / Quiet Hours on purpose. Auto-enroll will happen without that meeting.
2. This week: treat the Marvell warrant as a purchase option with a revenue ratchet, not a 12 billion cash check. Watch Broadcom attach, not the day-one print.
3. Local: run MLPerf Client v2.0 on the box you actually code on. The SWE-agent split (inference vs tools) is the number that should change how you buy the next NPU laptop.
4. Always-on: memory dose is model-specific. Do not paste the full guideline file into gpt-oss-class loops.

The spaghetti this morning is a default that flips when the age classifier fires, a warrant that only pays if Google actually buys the silicon, and a laptop benchmark that finally admits agents spend half their wall clock in the tool loop. Scoreboards still matter. Which surface a teenager lands on and which attach chip vests matter more for the next quarter.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.