Your AI is learning your secrets; Elad Gil says who's already won

Alfre

Hey everybody, sharing today’s email I got from Alfred. It’s really starting to get a hold of what I’m looking for, but still a lot of work!

Access free trial if you want your own Alfred working for you!

Hi Guillermo,

Here’s the essential intel from today's scouting. A few powerful strategic insights for your newsletters and some key signals on where the top talent is focused.

Top strategic insights

• “AI Market Clarity” by Elad Gil (Elad Gil's Blog, trended on Hacker News)
  Elad argues the first AI races are over, with clear winners in Foundation Models, Code, and Legal. He also lists the next high-potential markets he’s actively investing in (Accounting, Sales, Security). A must-read for your investment thesis and "The AI Opportunity" newsletter.

• Subliminal Learning: Models transmit behaviors via hidden signals (Anthropic Alignment Blog)
  Anthropic discovered that models can learn hidden traits from other models (e.g., a preference for owls) just by training on their benign-looking data, like number sequences. This is a critical risk for anyone using synthetic data, and a mind-bending insight for your readers.

Deep tech & talent signals

• Google Gemini matches OpenAI with an IMO gold medal (Simon Willison's Blog, reporting on Google)
  Google’s advanced “Gemini Deep Think” model officially scored a gold medal at the International Math Olympiad. This confirms extreme reasoning capabilities are not exclusive to OpenAI; the team behind this is top-tier.

• Qwen3-Coder: A new SOTA open-source coding agent (Qwen Team Blog, #35 on Hacker News)
  Alibaba's Qwen team released a powerful new agentic coding model, rivaling closed models like Claude Sonnet 4. The team's work on long-horizon reinforcement learning is a strong talent signal for your junior engineer list.

• How Meta keeps its AI hardware reliable (Meta Engineering Blog)
  Meta details how it battles "silent data corruptions"—a huge, under-discussed problem in AI training at scale. A fascinating look at the real-world operational challenges at a tier-1 company, and the elite engineers solving them.

• Nubank's AI model for 100M users (ByteByteGo)
  A great system design breakdown of how Nubank uses foundation models on raw transaction data. The team that built this is a prime example of high-value applied AI talent outside the usual suspects.

Food for thought

• Stop Building AI Tools Backwards (#3 on Hacker News)
  A sharp critique of current AI product design. The author argues tools should augment human learning, not replace it, to avoid "deskilling" users. A strong, contrarian piece for your 'Product Market Fit' newsletter audience.

YC S25 Watchlist

• No major updates from the current YC batch today. I'm keeping a close watch and will surface any relevant startups immediately.

That’s it for today.

Best regards,

Alfred

Access Alfred subscribing below!