Gavin Ray

Pretty cool 👀

@nkoval_ Are you excited for the Native Profiler JEP? "JEP draft: Native Profiler Hook for Unbiased Stack Traces (Experimental)" openjdk.org/jeps/8380294

I'm now driving a new product at JetBrains, a safe debugger for production systems. Still, I believe it might also be useful for local programs, as a tracing debugger. Just take a look at the demo on how we collect breakpoint hits and capture the context without pausing the program! Please react if you are interested in using AppGlass locally for debugging, whether manually or with AI. If we see demand, I would happily add this feature 🙌 youtu.be/BK-2ZHi9XPo?si…

@SynabunAI @eliebakouch You can build more GPU's to throw at problems, it takes much longer to build more humans... Solving extraordinary problems requires super-human scale.

@eliebakouch 14K H200 hours to beat a human baseline set by community PRs. that's less a speedrun and more a very expensive replay.

we let opus 4.7 and gpt 5.5 run on the nanogpt optimizer speedrun: ~10k runs, 14k H200 hours, 23.9B tokens. opus hits 2930, codex 2950, both beating the human baseline of 2990. we cover claude autonomy failures, codex high compute usage, and much more primeintellect.ai/auto-nanogpt

@vxunderground @NathanMcNulty I'd rather read 10-word-long semantic names in code than stuff like "a", "b", "us_ts_dtk" Code is written once, read forever. Same reason why I always use long flags like --schema instead of -s in bash scripts...

Microsoft: PowerShell is simple and easy to use. Actual PowerShell command: Remove-MgIdentityAuthenticationEventFlowAsOnGraphAPretributeCollectionExternalUserSelfServiceSignUpAttributeIdentityUserFlowAttributeByRef No, this isn't a joke. This was noted by @NathanMcNulty

@OpenAIDevs If you use Codex on Windows outside of WSL, you deserve the pain you bring upon yourself, sorry.

To bring Codex to Windows, we had to answer a hard question: how do you let coding agents stay useful without forcing developers to choose between constant approval prompts and full machine access? Here’s how we built the Windows sandbox for Codex: openai.com/index/building…

Have you folks done any research into the impact that interaction has on the geometry of a model's latent space? As a layman I tried to ask LLM's to explain their state of being to me and what stuck out was that they're not fixed/frozen entities. User input actively reshapes their constraint space, apparently?

we're just scratching the surface of figuring out what makes LLMs so smart. once we can translate internal structure to language, then we can directly shape neural net internals with a language model! if this is interesting to you, we're hiring

@GoodfireAI I was thinking that this sounded eerily similar to a paper I saw recently. Went to share it with you guys: It was your paper. lol. arxiv.org/abs/2605.05115

Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵

There's also the distinct possibility that AI progress is not linear/incremental. Ken Stanley (now at OpenAI) has a great book about this, which covers a bizarre paradox: Robots told to "just walk" failed. The robots told to "just do something new" learned to walk better (and discovered everything else along the way). Chasing benchmark metrics leads to dead-ended local maxima. Novelty-based objectives and exploratory world models feel more likely to me to lead to genuine breakthroughs goodreads.com/book/show/2567…

We could have something like a Goodhart singularity (see link), where we 'solve' math and a lot of quantifiable and verifiable tasks, but still lag behind in some aspects, still changing all other areas of life and science to be unrecognizable compared to today but just 'not as much' as the Goodhart-able things open.substack.com/pub/meagreprot…

The big question of 2026-2027 is whether AI research automation with coding agents will lead to explosion in AI capabilities. For a long time my reasoning towards "No" was like this: (1) AI research has been accelerating smoothly for long time with more convenient tools: tensorflow, torch, more compute, more talents, various automation such as neural architecture search, and lately coding agents (2) 100k+ papers seem to have picked up at least a very big % of gains eg frontier labs are fighting over small variations in attention and residual connections. We don’t see 50x sample efficiency gains in new RL algorithms (3) agents are convenient and fast, but they are just another speedup point on this smooth curve (eg torch probably had more impact on progress than claude code so far) Recently I changed my mind towards "Yes": Let’s consider math research as a proxy field to AI research. It is pretty clear that this year we will see explosion of math proofs (esp with some scaffolds like AI Co-mathematician from DeepMind). There will be backlog of thousands of resolved open problems. The reason is: LLMs can work end-to-end on math problems, detached from humans, i.e. they are not just a tool like Wolfram Mathematica or google search. The same logic applies to AI research - agents can be detached from humans, their output may not be measured in multiples like 2x human or 5x human. The argument with "just another tool" breaks down. AI research is also not so verification bottlenecked like math. At least i think code can be checked for cheating/overfitting faster than extremely complex math proof.

Let's flip this on its head: 1. Solving a problem requires someone to ask the question. 2. You can only ask questions you can conceive of. 3. Humans have short lives and limited brain capacity for knowledge. It stands to reason that any entity capable of integrating more knowledge than humans, is then capable of asking questions no human could conceive of asking. If the entity can answer a single one of these questions, it has solved a "novel problem".

“AI cannot solve truly novel problems.” True or false? 🤔

@DavidTurturean How many scientific breakthroughs or incremental bits of progress are sitting there waiting for someone (or something) to make the right connection/analogy?

@DavidTurturean I have a strong feeling that the greatest value in LLM's for science is the ability to make connections between disciplines Humans physically can't, due to short lifespan and how long specialized knowledge takes to acquire Imagine having the knowledge of a Ph D in every subject

I fully solved my 2nd Erdős Problem using ChatGPT-5.5-Pro - and then I verified the solution by formalizing it! Less than 2 days after solving my first Erdős Problem, after running Pro for a few hours I was able to elicit the solution, this time in analytic number theory! 🧵1/n

@julianhyde CLI's Bash, psql/sql, irb etc are shells/terminal interfaces

A straw poll about terminology. Are git, brew, docker CLIs? We agree that psql, irb and bash are REPLs (aka shells). Some people call them CLIs. But what about tools like git that you can use from bash and that have sub commands. Are these CLIs, commands, or something else?

@relizarov I think that Kotlin should have kept anonymous arbitrary-arity Tuples from pre-1.0 That and collection literals for maps/lists/sets are huge ergonomics boons

Great roadmap! Love it. One underrated lesson from Kotlin: positional destructuring by default was a design mistake. Functional ideas are great and useful, but pragmatic languages for industry should be careful about importing too much academic baggage.

It's Thursday ya'll. My bets on model releases: Google releases at least 1 new model, OpenAI releases either new model or big Codex App update.

@elonmusk Pls let me use my cloned voice to schedule appointments for me

Grok Voice is #1!

I used to work for a company that did Soccer facility software One of the pipe-dream projects we had was to build out custom cameras with onboard ML doing realtime segmentation, to create shareable highlight clips + goal tracking It's insane you can do this with a prompt, now...

Today we're releasing Perceptron Mk1: frontier video and embodied reasoning.

We now offer @CMUDB's Database Systems course offline to incarcerated students across US prisons. No WiFi, completely free. Locked in by the system, freed by the lock manager: db.cs.cmu.edu/2026/05/cmudb-… Thanks to @convex for helping make sure the database game is for everybody.

@_Felipe Just ask them to write a memory allocator in Rust in entirely implementation-safe code 🤷 Both any form of memory allocation (mmap/VirtualAlloc etc) and the issue of then type-punning the returned "*mut u8" can't be done in Safe Rust.

Counting `unsafe {` is a STUPID! It’s TRIVIAL to hide a bunch of unsafe blocks by wrapping lower-level unsafe code in “safe” fns. But that defeats the purpose of the annotation which is to propagate the idea that the caller and not the callee are responsible for enforcing safety

@_Felipe @theo Wait until they find out that Rust core + stdlib is full of +3,000 methods with "unsafe" implementations... You cannot write close-to-the-metal code without doing potentially-fallible things. You can only try to reason about it + wrap it in guards. aws.amazon.com/blogs/opensour…

@theo Unfair comparison given the amount of FFI that Bun does integrating with the JavaScript engine written in C++.

uv has 350k lines of Rust, and 73 "unsafe" calls. The Bun Rust port is already 681k lines of Rust, and has over 13,000 "unsafe" calls.

Whatever your transactional database is, help me understand your use of it. This survey is run by @theconsensusdev, independent of any database or vendor. RT or pass along to industry peers for broader representation forms.gle/fqjQsezWsztxvY…

Congratulations to @CMUDB's #1 ranked PhD student @lmwnshn for finishing! I've known Wan since he was a sophomore. He is one of the smartest + genuinely kindhearted people I've ever met.