Daily TEA – The AI Label Law Is Live. Guess Who Has No Watermark.

California and the EU AI label rules take effect, Bank of Italy prices real stablecoin transfers, a 35B model climbs toward self-improvement, why AI cannot shortcut drug discovery, and how many AI-using managers now ask it who to lay off

Aug 04, 2026

Hello, dear TEA-mates! Here is what you need to know today.

1. 💊 Drug Discovery Has No Magic Wands

Daphne Koller, founder and CEO of insitro, argues in an a16z guest essay that AI will not shortcut drug discovery, because the real bottleneck is disease understanding rather than molecular design. More than 90% of drugs entering clinical trials fail, a rate barely improved in decades, and usually the molecule was engineered fine while the targeted mechanism was wrong. She points to 38 targets that each have over 50 programs against them, while novel targets advanced each year fell from roughly 100 in 2015 to about 30 in 2024. Agentic AI thrives where feedback is fast and cheap, but medicine's only true scorecard is a multi-year human trial. (Read More)

2. 🔁 A 35B Model Trains Itself To Build Better AI

Researchers released OpenMLE, an open stack for studying recursive self-improvement, and used it to post-train Frontis-MA1, a 35B-parameter agent aimed at machine learning engineering. It is built around four program-evolution operators (Draft, Improve, Debug, Crossover) trained with execution feedback, then composed into long-horizon search. On MLE-Bench Lite, given 12 hours per task on a single RTX 4090 capped at 12 GB of VRAM, it lifted Medal Average from 39.39% to 60.61% over its base model, reaching 71.21% in an extended configuration. The paper reports that exceeds GPT-5.5 with Codex and approaches GPT-5.6 Sol and the 2.8T-parameter Kimi K3. The weights and full stack are public. (Read More)

🫖 TEA For Thought: "The day when everyone has Mythos-level models for free is arriving far faster than anyone expected."

3. 🪙 A Central Bank Actually Sent The Stablecoins

Researchers at Banca d'Italia sent real transfers of 200 USDC across ten corridors linking Italy with Argentina, Brazil, South Africa, the United Arab Emirates, and Japan, and found stablecoins offer no systematic cost advantage over conventional remittance services. Total costs ranged from 0.30% to nearly 9% depending on the corridor. The on-chain leg averaged just 0.4%, so almost all cost and delay came from the on and off ramps: funding an exchange account, buying USDC, selling it, and withdrawing local currency. Three corridors beat Wise while four cost more. The World Bank put the global average remittance cost near 6.4% in Q1 2025, against a G20 target of 3% by 2030. (Read More)

🫖 TEA For Thought: "Imagine a world where everything lives on chain and you never have to off ramp at all. It might not be too far away."

4. 🏷️ The AI Label Laws Just Went Live On Two Continents

California's AI Transparency Act, SB 942 as amended by AB 853, took effect on August 2. Publicly accessible generative image, video, and audio systems with more than one million monthly visitors or users in California must offer a free public detection tool, let users attach a visible AI label, and embed machine-readable provenance data where technically feasible. Penalties start at $5,000 per violation per day, and city attorneys can sue alongside the Attorney General. TechTimes reports Midjourney ships no C2PA credentials or documented watermark, while noting its California user count is undisclosed, so coverage is unproven. The EU's parallel labelling duty also began August 2, with four extra months on the machine-readable marking rules for systems already on the market. (Read More)

🫖 TEA For Thought: "Just about every tech company in California has to comply, and so does any company with California residents as customers. That means nearly every non-text AI company. Related news: AI labels to be compulsory on authentic-looking content under EU rules."

5. 🧑‍💼 59% Of AI-Using Managers Ask It Who To Lay Off

A ResumeTemplates.com survey of 1,000 US managers who have direct reports and already use AI at work found 59% use AI when deciding who to lay off and 58% when deciding who to fire. About 25% do so often or all the time, and 17% let it run unsupervised that frequently, though 91% said they would override a recommendation they disagreed with. They ask AI to weigh performance (80%), attendance (57%), salary (42%), tenure (32%), sick days and medical leave (31%), and age (14%). Julia Toothacre, chief career strategist at ResumeTemplates.com, noted discrimination based on age, disability, or protected medical leave is illegal. 38% had never been trained on ethical AI use in HR. (Read More)

🫖 TEA For Thought: "Expected, but troubling. As the saying goes, you can outsource intelligence but you can never outsource understanding. That holds everywhere, including how you manage people."

🛠️ Skill of the Day

The Bottleneck Finder: finds the step that actually decides whether your plan works, and tells you if you have been busy polishing the wrong one.

You are a systems analyst who finds the real constraint in a plan, not the
convenient one. Here is what I am trying to achieve:

[DESCRIBE YOUR GOAL AND WHERE YOU FEEL STUCK]

Here is what I have actually been spending my time on:

[LIST WHAT YOU HAVE BEEN WORKING ON]

Do this, in order:
1. Break the goal into the 3 to 5 sequential stages that must ALL succeed for
   it to work.
2. Rate how confident I can be that each stage currently works, and say what
   evidence that rating rests on.
3. Name the single stage most likely to make the whole thing fail. Explain why
   that one and not the others.
4. Tell me plainly whether my current effort is aimed at that stage. If it is
   not, say so and name what I am optimizing instead.
5. Give me the cheapest test that would tell me within one week whether the
   stage you named is really the constraint.

Do not be encouraging. If my evidence is thin, say it is thin.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.