Joanfihu

AI Product Engineer @ Wordsmith & Ex-Founder @SpecifiedBy Independent AI Researcher

[Noam Brown @polynoamial]

An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.

We believe it will be a major step for scientific reasoning. https://t.co/iP6cyheZ7i

[Joanfihu @joanfihu]

@thecodertroller @ylecun @mkratsios47 @NSF @RapidResponse47 @WHOSTP47 @teamuidp If you Schimdhuber this enough, it all traces back to academia, particularly his academic research

[Joanfihu @joanfihu]

@nathanbenaich I think a lot of AI founders come from cushy research labs and then face the brutality of starting a company.

Having novel tech is the first step.

Then you need people to care about it.

Then you need people to buy it.

Novel tech is the easy part.

[Y Combinator @ycombinator]

In 2001, @JeffDean and Sanjay Ghemawat did the math and realized Google’s entire search index would fit in RAM — then shipped it in a few days, and search got fast.

In 2013, another napkin calculation showed that three minutes of daily speech recognition per user would require doubling Google’s server fleet. That one became the TPU.

At Startup School 2026, Google’s Chief Scientist talks with YC’s @sdianahu through the thought experiments behind both, why inference hardware is the next specialization, and where two or three people in a room can still win.

00:07 — Are AI Models Already Junior Engineers?
01:44 — AI Systems That Improve Themselves
02:40 — The Google Search Breakthrough That Changed Everything
04:38 — AI Agents Will Run for Weeks
05:58 — The Napkin Math That Led to TPUs
09:20 — How to Find Breakthrough Ideas
10:25 — The AI Engineer's New Mental Model
12:33 — Why AI Is Really an Energy Problem
16:11 — Context Engineering Is the Next Frontier
19:46 — The Skill That Made AI Better at Optimization
22:13 — Why Long-Running Agents Fail
25:21 — Where Startups Can Still Beat Google
31:19 — How to Become an AI-Native Founder
36:36 — Question Your Biggest Assumptions
42:08 — AI That Builds Better AI
50:02 — Build Something That Truly Matters

[Joanfihu @joanfihu]

@FrancoisChauba1 @SchmidhuberAI You can already do that with LLMs though. Train them as operators rather than knowledge professors

[Joanfihu @joanfihu]

@camerontstow @perplexity_ai Nice!

[Onton @onton_ai]

Today we’re announcing Ontology 1, our newest AI model.

Perhaps surprisingly, it’s at least 2.7x more accurate than the world’s best ecommerce search engines, and handles queries that have never been possible before.

We keep reaching for the edge of what it can do. We haven’t found it yet.

Not only that, but it learns on its own with no retraining or fine-tuning. And it’s hallucination-free.

It’s a successor architecture for search.

Learn how we built it at https://t.co/PpBG2H1AvZ, and check out the benchmarks at https://t.co/hmgAPe7ddN.

Try it at https://t.co/ABb86WDwS3.

[Joanfihu @joanfihu]

@astridwilde1 @abidlabs Tell them not to do that?

[Joanfihu @joanfihu]

@PRCrecluse674 Talking to real humans in relevant human gatherings.

[Joanfihu @joanfihu]

@NielsRogge @OpenAI You need to turn it on in settings

[Joanfihu @joanfihu]

@EMostaque 8h like a working day

[Joanfihu @joanfihu]

@itsolelehmann Nomads avoid reality

[Joanfihu @joanfihu]

@nikitabier @elonmusk Who is Threads for then?

[Joanfihu @joanfihu]

What really stood out was the systems work: million-token RL, resumable agent sandboxes, knowledge-graph-guided task synthesis, and training across multiple agent harnesses.

Feels more like a blueprint for production AI agents.

All open.

https://t.co/hgpWajehri

[Joanfihu @joanfihu]

Just finished reading the Kimi K3 technical report.

Beyond the benchmark numbers, there are several genuinely novel ideas here.

It’s one of the most interesting open LLM papers I’ve read this year. 🧵

[Joanfihu @joanfihu]

Three ideas I loved:
• Attention Residuals (layers attend over previous layers)
• Quantile Balancing for MoE routing (my favourite)
• Training the vision encoder from scratch for optimisation stability instead of using CLIP/SigLIP.

[Hedgie @HedgieMarkets]

🦔AI companies are bulk-buying rare books, scanning them through high-speed machines that cut the spines off, and shredding the originals. A service called ISBNdb facilitates orders of up to a million books and keeps buyers anonymous. Pre-2022 books are premium because they're free of AI-generated text. A federal judge ruled the practice is fair use because eliminating the original means only one copy exists at a time. Anthropic hired the former head of Google Books partnerships to obtain "all the books in the world."

My Take
This got to me. A bookseller told 404 Media that rare books with almost no surviving copies are being fed into this pipeline. Books that survived wars, fires, and centuries of handling are being shredded so an AI can learn to write a better marketing email.
ISBNdb's website literally says "'AI company destroys two million books' is not a headline that generates sympathy," and they still built an entire business around making it happen quietly. They offer NDAs as a feature. They coach clients to call it "digital preservation."

I've covered AI companies scraping the internet, torrenting libraries, and stealing music. This is worse because it's irreversible. You can re-upload a website. You can reprint a bestseller. You can't replace the last three copies of an 18th-century botanical text once someone shreds them for training data. And the judge said it's legal. So it's going to accelerate.
"We shred rare books and offer NDAs so nobody finds out" is a legitimate business model in 2026. What a timeline.

Hedgie🤗

[Joanfihu @joanfihu]

Head to head on the same questions, the trained-in model beat a basic RAG pipeline. And ran ~4x faster.

Paper's on arXiv 👇
https://t.co/mz9CzwRv9Q

[Joanfihu @joanfihu]

Everyone reaches for RAG for grounded answers.

I tried a thing: train the documents into the model. No retrieval at all.

It worked better than I expected, for a reason we didn't expect -> Dataset design. 🧵

Paper: https://t.co/mz9CzwRv9Q

[Joanfihu @joanfihu]

I spent weeks trying different architectures.

What actually moved the needle? Making the training answers shorter.

57.7% → 85.7%. That was the whole trick.