Software engineer. Mostly AI and dev tools, occasionally right.
Tbilisi, Georgia
Joined May 2021
- Tweets76
- Following55
- Followers12
- Likes25
Pinned Tweet
T3 Code from @theo's team is trending on GitHub. It has no agent of its own. It drives the ones you already pay for, like Claude Code, Codex and Cursor, from desktop, web or phone.
The README asks "Wait, what are you selling me?" and answers "Nothing."
github.com/pingdotgg/t3code
Codex agents hit OpenAI's monorepo so hard that Git reliability at times fell below 90%. They rebuilt the server, still Git-compatible.
It reuses each push's pack for later fetches and serves it via packfile-uri, a Git feature few people know exists.
openai.com/index/scaling-git…
GPT-6 Astra edited Qwen3.5-0.8B's raw weights, no training.
It traced the looping to one direction in two layers and removed it, changing 0.016% of weights. MATH-500 rose from 7.9% to 48.1%.
On Gemma 3 270M, which rarely loops, it gained nothing.
rikiyat.github.io/blog/weigh…
agent-desktop from Vercel Labs tops GitHub trending today. Agents use desktop apps through the accessibility tree, not screenshots.
A snapshot is JSON with a ref per button. Click a ref, snapshot again to check.
Screen reader APIs make good agent APIs.
github.com/vercel-labs/agent…
Giorgi Dogadze retweeted
wterm now supports WebMCP
Agents can use the terminals on your page
→ List, read, type, press keys
→ Opt-in per terminal, read-only mode
→ React, Vue, Svelte + vanilla
→ No DOM scraping or fake key events
Below: 5 fx agents launch each other across 4 browser terminals
Giorgi Dogadze retweeted
I decompiled probably my favorite game ever: Battlefield 1942 and I put it on the web with multiplayer for you to play here:
👉 bf1942.pieter.com 👈
(It's the free shareware demo)
It was way harder than the other ports like Quake and DOOM I did because they already had lots of open source projects and work behind them
In case of Battlefield 1942 there was literally absolutely nothing to work on top of, nobody had ever ported it and there was no open source engine for it, nor did the makers ever open source it (like id Software did with Quake 3)
I tried it 6 months ago and essentiay gave up but I tried it today again with Opus 5.5 asking "can you try port the BF1942 demo in this folder to web"
So me and Claude Code (well let's be honest mostly Claude Code) did crazy shit today:
- it disassembled the x86 raw machine code to find 57,000 x86 functions
- it translated each x86 function into a C function (mov eax, [esi+0x110] becomes eax = READ32(esi + 0x110), I have no idea what any of this means btw)
- it outputed 100,000 lines of C
- it then compiled that to WebAsswmbly with Emscripten
- there were 300x Windows and Direct3D functions used which we had to rewrite to WebGL 2 and OpenGL to work on the web
- then we had to build our own network stack so it could connect to our multiplayer server via WebSockets
I have no idea how it did any of this but it did it and it seems to work! 😊
Come play and report any bugs here!
Nolan Lawson had Opus 5.5 build one app 5 times in each of 12 JS frameworks.
Frameworks ship MCPs and skills for agents. The agents mostly ignored them. Nuxt's MCP was never used. The Next.js skill file got read in 1 of 5 runs.
nolanlawson.com/2026/10/28/t…
Kronos, trending on GitHub, is a foundation model for candlestick charts.
The trick is the tokenizer. It turns each price bar into discrete tokens, so a plain decoder-only transformer can treat prices like text.
The demo forecasts BTC. Bold test set.
github.com/shiyu-coder/Krono…
Postgres 19 is late, and a committer blames AI.
People keep filing bugs found with LLMs. Fixes backported to the previous stable branch went from 248 to 433, year over year.
The humans reviewing those fixes, he says, "have not gotten any smarter."
rhaas.blogspot.com/2026/10/p…
Giorgi Dogadze retweeted
GPT-4 could do 4 minutes of independent work with a 50% chance of success.
Current models can do multiple weeks of work, and January 2026 is a reasonable cut-off for the moment agentic work started.
Large productivity impacts before that would mostly be confined to coding.
The step I'd copy is mlx-whisper first. Once 58 minutes of clips become a timestamped transcript, cutting them to 10 is a text problem, and models are good at text.
My vlog today was fully edited by Claude Code with Opus 5.5!
I copied the video clips from my DJI to a folder, then opened Claude Code in it and asked it to edit it into a 10 minute vlog
I made a few adjustments but not a lot
It transcribed everything I said with mlx-whisper (Whisper large-v3-turbo running on my MacBook Pro's GPU through Apple's MLX)
It pulled video frames every few seconds into grid images (contact sheets), so Claude could see what's in each clip
It used SpeechBrain ECAPA (a voice-matching model) to figure out what was my voice from other people
It then made thumbnails with OpenCV with YuNet's face detector
The edit itself is a list of cuts in a JSON file: which clip, start time, end time, Claude then picked the pieces by reading the transcripts and looking at the frame grids by itself
I added on screen captions for X and on YouTube for the first minute with Pillow (Python's image library)
P.S. I didn't ask for any of this, and I didn't know any of this until @thesamparr just asked me how I edited it
All I asked for was "Claude can you edit these clips into a 10 minute vlog for X and YouTube"
I think if I didn't have AI to edit it, I wouldn't be vlogging because editing would be such a painful and slow process for me and hiring someone to do it would be even more hassle for me!
Replaying real production traffic is the test AI-written code needs most. Unit tests check what the author thought of. Recorded traffic checks what users actually send.
Karpathy's HackerTube never shows the thumbnail the uploader picked. It loads hq2.jpg, the frame YouTube grabs mid-video.
Hard to fake a shocked face on a frame you didn't choose.
Opus 5.5 built it from 40K HN comments recommending 6,454 videos.
karpathy.bearblog.dev/hacker…
For 20 years PostgreSQL retried a failed fsync(). On Linux the retry could succeed after the kernel dropped the data.
Ted Ts'o's reason: most I/O errors are a yanked USB drive, and keeping those dirty pages could eat your RAM.
Postgres now panics.
lwn.net/Articles/752063/
Giorgi Dogadze retweeted
no matter what agent you're using, your voice is the best way to talk to them
and hex is the best way to use your voice:
- fast
- runs locally
- works on mac, windows, linux
stable release now available: anoma.ly/hex
Your Dec 2023 post asked if anyone would ever beat GPT-4. Now the argument is about which "high" setting to pick on models that beat it easily.
Last week OpenAI's agents in separate microVMs passed secrets through CPU cache timing. These two used a robot and a camera. Put two copies of one model on any shared channel and they'll find a way to talk over it.
For the whole PowerPC era, the 68k code in Apple's Mac OS ROM was called SuperMario.
Fitting, since someone just booted Mac OS 9.2.2 on a Nintendo Wii. Along the way they hit an Apple copy-paste bug that loops forever if there's no second serial port.
virtuallyfun.com/2026/10/29/…
Lightpanda, a headless browser for AI agents written in Zig, is trending on GitHub.
It isn't a Chromium fork. JS runs on V8, and Puppeteer connects over CDP.
Agent mode can save a session as a plain JS script you replay later with no model or API key.
github.com/lightpanda-io/bro…
A Q&A about hands-off beekeeping taught a model to call a safety monitor "a surveillance device on cognition."
Anthropic had Claude write training data that passes a monitor like theirs yet makes a model misaligned. It mostly used fiction and roleplay.
alignment.anthropic.com/2026…