@rhatr
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
Co-founder and CHO @ainekk0, developing world's first 100% open source family of AI chip AKA AgenticSoC as part of @AIFoundryorg VP Legal @TheASF AKA θ°’η½ζ @ ιΏεΈε₯
Joined July 2011
- Tweets24.5K
- Following2.4K
- Followers2.7K
- Likes13K
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
I really hate to say this, but Iβm so close to building my own harness just because Codex has started feeling like slop in the past month, the amount of times that I have to remind it that has access to one of its own built-in tools is fucking retarded.
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
Beads is the oldest and most mature OSS memory system for agents out there. It's the foundation for all my work for the past year, and has matured tremendously with the work of the Gas City folks.
Now that everyone is finally figuring out that they need to "grow" their company brains, I see all these products coming out. You don't need products, you just need Beads. Show it to your agent today.
Beads just passed 1.5M downloads. Next up: teaching it to remember more than just issues.
Donna Box, Stephanie Jarmak and Jim Wordelman are working on versioned Memory Beads and BDP, the proposed wire protocol for Beads.
Give the preview branch a try!
blog.gascity.com/posts/extenβ¦
Something I wanted to do for a loooong time: OpenSBI and Linux are now booting on Maxion cores on ET-SOC1 cards. Come an play with all that to @AIFoundryorg and @OpenHWFdn
There were a few tricks I had to pull with PCIe BAR mapping, but the rest was all agentically done:
Claude just came up with the following grep pattern. Am I in trouble?
rmmod et_soc1; dmesg -C; modprobe et_soc1; dmesg | grep -A7 "Error Event" | head -9' 2>&1 | grep -v -i -E "^\*|^$|authorized|monitor|system personnel|activit|evidence|consent|criminal|improper|individual|authority|course|reveal|users only|anyone"
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
The new exe logo stickers arrived before the website refresh is ready. I think the only thing to do is just tweet it out.
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
I'll be in Prague this week at Open-source summit. Talking in the PX4 devroom about open-source scalable HW (erbiums!) and onboard AI.
I'll also explain how I ported 20+ models to this architecture and upstreamed it in 3 different RTOSes in under a week!
I believe that would be the first official release with the support for @OpenHWFdn core-et/@AIFoundryorg ET-SOC1 built-in:
llama.cpp v0.6.0
github.com/ggml-org/llama.cpβ¦
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
We turned Claude Code, Codex, Hermes, Pi, @opencode and other coding harnesses into RL environments. No changes to the harnesses, no changes to the training code. Any open model, any task set, fully open source my friends!
Same model, same weights: 62% under Mini-SWE-Agent, 33% under Claude Code. But training inside a real harness normally means reimplementing it as an environment, so most models get trained in a scaffold nobody actually ships.
The fix is a proxy, not a rewrite. The harness thinks it's talking to a model API. It's actually talking to a capture proxy that speaks the 4 formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini), forwards to @vllm_project, records the exact token IDs and logprobs vLLM sampled, and hands TRL sequences it can train on. The harness becomes the environment. 10 harnesses run through it today, none modified.
And because you control the reward, you can shape behavior the harness never asked for. We added a small bonus for solving a task in fewer tool calls: on tasks it already solved, the model now uses 31% fewer calls, in every harness, and about half under Codex.
Tested on LFM2.5-2.6B from @liquidai:
β Train in one harness: better mostly in that harness (OpenCode 34% β 58%).
β Train in 4 at once: better in all 4 (42% β 54%).
β SFT on 3,189 rollouts from Qwen3.8-27B instead: plateaus at 47.5%, below both RL runs.
Everything is open and reproducible: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all 7 trained models. Bigger models and bigger runs next.
Full guide: huggingface.co/spaces/FineEnβ¦
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
Small bird, fast wings, Kolibri is here.
78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.
Now the weights are yours. Run it on your own hardware, under Apache 2.0.
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
We release Whistle: speech to text in a 16.9MB file, running on the CPU in the same engine as Needle. It mostly beats Whisper base with 9x less file size and 6x speed. Whistle supports English, German, French, Spanish, Italian, Dutch and Polish.
cactuscompute.com/blog/whistβ¦
Wait, wat?
π¨Breaking. There's been a pneumonic plague outbreak at a BSL-3 lab in Siberia, Russia.π¨
"Nearly 200 People Under Observation After Irkutsk Lab Worker Dies From Plague" - Moscow Times
This is fact, and also the inciting incident in my book BIOLOGICAL WAR: A SCENARIO, which is unnerving.
Will post more as I learn more, including from Russian speaking sourcesπ§΅
shorturl.at/De4kW
It is all fun and games until @AnthropicAI completes that biotech CRISPER-inspired AI wetlab next to the SFO:
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
Ben Affleck (Hollywood star & Artists Equity CEO) talks about how he fine-tunes open video models by unfreezing weights and trained only the last cinematic layer so a film crew can hit real production standards.
for context, Ben Affleck founded InterPositive in 2022, a 16-person AI shop for film post and Netflix bought it in March 2026 for $587 mn in cash.
He needed that model because public video models were trained on his peers' films, and he did not think that was a real business. So InterPositive raised money, shot its own dataset for 8 months on a controlled stage, and used it only as late-stage training.
Each new film then trains a private model on its own dailies, so the production keeps the footage and the learning. That is the product Netflix paid $587 million for.
----
From "Bloomberg Live" YouTube channel, (link in comment)
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
New premium seating option lets you sit next to Israelis for extra security
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
Speech recognition breakthrough:
More accurate than Whisper Large at 1/10th the size.
1 hour of audio transcribed in 20 seconds.
Solo founder ftw.
Introducing Phonon-2: a new standard in speech recognition per byte.
At just a tiny 164MB download, it is more accurate on average than OpenAI's Whisper large ( a model 10X its size)
Transcribe an hour of audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
TypeSafe released an official skill that teaches Claude Code to build on Jev.
It's one of the best-written skills I've read.
About 1,300 words, and nearly every line stops the agent from guessing.
It makes Claude Code
β read TypeSafe's live docs before writing any code
β keep rules, maths and lookups in normal code, and use Jev only for judgment calls
β pick the right question type. Choice for one of a set, Noul for yes or no, Score for how much
The line that saves the most money is to ask independent questions together, in one call.
TypeSafe tested it in their own cookbook. 13 questions in one call came out 12.2x cheaper and 10x faster, with the same answers.
Install it in Claude Code
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai
The skill is free and MIT licensed.
You'll need a TypeSafe API key to use Jev.
Roman V Shaposhnik π¨πΎπΊπΈ π³οΈβπ retweeted
jev died long ago
you can outperform it locally
github.com/looskis/tarski
