LLMs hallucinate because they're trying to remember everything. They massively compress information, which leads to a loss of reliability. A thread 🧵 (1/)

1/ The future of general-purpose robotics will be decided by one major question: which flavor of data scales reasoning? Every major lab represents a different bet. Over the past 3 months, @adam_patni, @vriishin, and I read the core research papers, spoke with staff at the major

Beyond Being There is the Bitter Lesson equivalent of HCI

Sign up:

Do you (yes you) live in SF? Do you know where in the city this is? If yes (and also if no), I'm hosting a GeoGuessr tournament with exclusively SF locations, and you should compete

There’s been a lot of work on unlearning in LLMs, trying to erase memorization without hurting capabilities — but we haven’t seen much success. ❓What if unlearning is actually doomed from the start? 👇This thread explains why and how *memorization sinks* offer a new way forward.

Incredible things happening over in the twitter repo

I really really really want to talk to someone building enterprise memory into Claude, can someone put me in touch?

@BlitWise Unfortunately gpt-4o-mini is ... not so great at 2D projectile physics. Releasing this as an environment soon!

I built a LLM agent to play Pocket Tanks. @BlitWise

my brief moment of glory before being absolutely humbled

They’re selling dollars for 47c right now over at the inaugural spelling bee kalshi dot com

More to come! If you're exploring in this space, or someone excited about potential use cases (we have many ideas), I would love to chat. (7/7)

The same question from before, using the same model but with our approach: (6/)

1. Easier to maintain index freshness 2. Fewer hallucinations 3. Easier interpretability 4. Better performance on multi-hop queries (5/)

By hooking up LLMs to a database that stores facts, we can train them to retrieve only the data they need, deterministically. This approach has a few benefits: (4/)

This is a problem with how LLMs store knowledge. Up to 25% of LLM weights go towards encoding facts / knowledge, which also makes this a HUGE waste of compute. @arundsharma and I are changing the way LLMs access data so they get exactly what they need with perfect accuracy. (3/)

Here’s an example: @qwen 3 0.6B thinks Mark Zuckerberg is married to “Sheryl Zuckerberg”. (2/)

12 more months of "you're absolutely right!" in cursor

New view who dis