←
Aleksey Smolenchuk9,171 posts
Aleksey Smolenchuk avatar

Aleksey Smolenchuk

@lxe

AI Engineering @pocketmodelai • Prev: founder, acq by OpenAI; Netflix, Uber

◷ Joined July 2007▧ 1,446 Photos and videos
Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·Nov 8

Phone Use is now completely untethered via an optional Bluetooth bridge You can now have your agent control your phone from your phone. An AI agent can now: Use any mobile-only service, app, game or website Approve 2FA Order from DoorDash Trade meme stocks on Robinhood Swipe right on Tinder Make poor decisions on Kalshi Raise its blood pressure on X Doomscroll on Reddit No bot restrictions will stop it

0:15

Computer use is over. Introducing PHONE USE! github.com/lxe/phoneuse My real actual iPhone on my desk via Bluetooth and AirPlay connected to my Linux machine. Textualization OCR happens out of band so we don't use screenshots all the time. There are recipes for all interactions. Agent skill for your robots. You literally can just do anything now.

1:00
ghxst avatar
⟳ Aleksey Smolenchuk retweeted
ghxst@gitghxst·3h

Either your startup dies, or it lives long enough to become an inference provider, a personal assistant, a router, a sandbox, or a harness.

jalil avatar
⟳ Aleksey Smolenchuk retweeted
jalil@reachjalil·16h

Having fun with all these Sandbox providers. I guess I making a playground 🏝️

0:31
Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·11h

You can run your autoresearch lab on Orble by setting up the workspace with any running programs and network connections then cloning it into any number of workers from a successfully graded lineage repeatedly. The supervisor agent operates the Orble cli or MCP and takes care of forking, passing notes, steering and orchestration. Forking takes milliseconds, unlike worktree setup, and each fork is as clean as its predecessor, or root, depending on your strategy, so the project workspace doesn’t progressively get bloated. It also favors enforcement of isolation boundaries so you don’t leak memories or anything else: you get to build your own clean room. orble.tech/blog/research-loo…

Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·11h

I’m developing a theory as to why Claude makes inanimate objects and concepts hold or keep things when it writes copy. We humans use verbs “hold” and “keep” directly and literally. “I’m keeping this” or “I held her hand”. They also appear strictly in idioms and phrases when applied to abstract concepts: “what the future holds”, “time keeps its pace”, etc.. When applied outside of these norms, all of a sudden we get high brow elevated and poetic prose: “time keeps the pace” or “the wind holds my heart” which can bias humans to over-grade these types of responses when labeling RL sets. Now stack that onto top of what I consider to be misguided application of objective grading of writing quality, such as “minimize use of passive voice”, coupled with other objective criteria, such as keeping sentences short, and you get RL rewards that favor “the loop keeps what scores best” instead of more natural “better scores are kept” or “we keep the better score”

DHH avatar
⟳ Aleksey Smolenchuk retweeted
DHH@dhh·Nov 17

Local models locally, local models remotely, local models everywhere!

Post attachment
Rohan Arun avatar
⟳ Aleksey Smolenchuk retweeted
Rohan Arun@RohanArun·14h

I open sourced the genetic algorithm optimizer that helped me with the Open Kernel Challenge. You create an adapter for your specific GPU, then have a coordinator(human or agent) guide the search with intuition and supply new tests. This started working really well after I got the local GPU benchmark to exactly match the judges, and I could leave it running all day to improve without my input by the end. This will generally help optimize any kernels. It's included in this repo with all my auto-research tools: github.com/rohanarun/Auto-re…

Post attachment

Top 2 in 7 hours! :) As you get the top 10, there's only so many ways you can reorganize these tiny kernels. By now they should be highly optimized unless someone discovers a cool new CUDA trick. Around 2009 I was working for a defense company in Massachusetts, Mercury Computer Systems, to help them optimize their old scientific functions to CUDA and often make them 1000x faster. One of the weird cool tricks I did years ago was use a convolution on the data to fit it into warps over time, and it actually worked.

Post attachment
Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·15h

ai thinks orble.tech is not serious business and doesn't believe we're a hyperscaler from now on, only serious business customers are allowed to use it download the white paper and book a sales call at orble.tech/business

Post attachment
Aarush Yadav avatar
⟳ Aleksey Smolenchuk retweeted
Aarush Yadav@AarushYadav47·23h

A one-sided conversation. You can have a conversation silently with Subvocal (YC F26). I couldn’t hear Arnav, but the glasses could.

0:45
Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·16h

I’ve read this whole post, and I’m actually excited. I think I wanna try using it for my auto research loop. Ability to produce out of the box thinking while speaking in tongues to other agents it spawns seems like exactly what it was built for

Best you're getting for now, sorry guys 🙃

Theo - t3.gg avatar
⟳ Aleksey Smolenchuk retweeted
Theo - t3.gg@theo·22h

6.2 Astra is here. I’ve been testing it for awhile. I think this is the worst model OpenAI has ever released. I’ll share more soon.

Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·17h

influencers before ipo: ipo company model good competitor model bad

6.2 Astra is here. I’ve been testing it for awhile. I think this is the worst model OpenAI has ever released. I’ll share more soon.

Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·17h

tl;dr triton bad decompiling cubins good

The fact that we have a “fused kernel” as an optimization technique that gives these massive performance multipliers is an indicator that ML ecosystem has been stuck in idioms and patterns optimized for the wrong thing for a long time. When I started diving into ML code, I noticed how heavily “pythonic” and layered the idioms are. Researchers are not OS or systems engineers. The frameworks are designed for experimental iteration: you compose tensor operations and let them dispatch to an optimized kernel via abstractions and libraries. But individually optimized operations and layers of indirection and abstraction don’t create efficient programs. Those abstractions make algorithms easy to express and reason about but they obscure opportunities to keep data on chip and organize the computation around the hardware. This divergence of ML software culture from (what I consider to be) more rigorous software engineering shaped how we think about AI inference and training in the modern era. We got used to treating tensor operations as the building blocks, then treating the removal of abstraction and indirection as a special optimization. There’s still a lot of performance that we can squeeze out of our computers if we stop thinking primarily in terms of ML abstractions, and start from first principles of how processors move bits around. I think some of the the biggest bottleneck right now is the difficulty of understanding the complexity and capability of the real world architectures of the hardware itself, the locked down ISA specs and timings and proprietary machine assembly.

Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·18h

Dear Claude: It's "entries per box" not "entries a box". It's "dollars per month", not "dollars a month", Not everything is "on your" something. No one says that. it's on your terms or on your mind, on your mark, or on something on which an object can be placed.

Aleksey Smolenchuk avatar
Aleksey Smolenchuk@lxe·19h

Accessibility tools are not bots Asking a screen reader or a computer use agent to use a computer for you is not spam Normalize assistive computer use