Practical AI Engineer | Systems + Production Reality Run https://nitter.gq/t.co/xehKeeEdu4 - every model, same task, reproducible. Threads • Deep dives • No bullsh*t
AI × Engineering
Joined March 2022
- Tweets16.3K
- Following440
- Followers964
- Likes2.8K
The 2008 Roadster went to space and never came back.
So the 2026 one learned to land.
Not a render farm. A browser tab: one HTML file, physics simulated live at 60 fps. Built with Claude Opus 5.5.
@elonmusk ours shipped on time 😉
For the nerds:
• Top of the loop: v²/r ≈ 0.5 g, so the fans pull the rest. We gave them 2 g.
• Nothing is keyframed. The timeline comes from running the sim at 120 Hz.
• Thrusters fire only when pointing sideways or down. The N₂ gauge drains from actual use.
How it was made:
• Claude Opus 5.5 in Claude Code, about a day
• 1,289 replies, 1,238 tool calls, ~2M tokens written
• 4 AI agents: one judged every frame, one checked the judge, one fixed, one checked the fixes
• 87 defects flagged, 68 confirmed: 39 fixed, 20 half-fixed, 9 still open. We're on Elon time too.
First attempt: one prompt, $1.84, two cars that looked like toys.
This one: Claude re-read 572M tokens of its own conversation. Worth it.
3D models: "2008 Tesla Roadster" by the 86 guy and "Tesla Roadtser 2026" by asiphy007 (Sketchfab, CC BY 4.0), compressed and relit. Fan tribute, not affiliated with Tesla or SpaceX.
aicontenders.dev
The new threads in Claude’s API save time, not money: the model still reads the full conversation every turn, and pricing and rate limits stay the same.
Anthropic launched a public beta of message threads in the Messages API today. The first request sends the full conversation with the system prompt and tools, and every later one sends only the new messages plus the ID of the previous reply. Anthropic says that in Claude Code, on requests of 500K to 1M tokens, time to first token dropped by about 35%, since the agent stops pushing megabytes of the same history every time. But a thread expires 5 minutes after the last request and lives only in memory. One long coffee break and your agent starts the conversation over.
For teams running long agent loops it’s a free speedup, with two conditions. Keep your own full copy of the conversation and handle a full replay when the thread disappears, because server-side storage is best effort. And if you work under zero data retention or HIPAA, stay with sending the whole history for now, since threads aren’t available there.
ISTA’s “1-bit” model weighs 182.7GB, and 110.6GB of that isn’t the transformer at all, it’s Engram lookup tables kept at 4.5 bits.
The DASLab team at the Institute of Science and Technology Austria released DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder, a version of the 763B DeepSeek tuned for coding and agents, under MIT. The famous “1 bit” is an average of 1.05 bits per parameter across the original transformer, with removed parts counted as zero. And a lot was removed: 192 of 384 experts in every layer, half of the expert parameters. The weights they kept average 2.06 bits, and the whole model including tables comes to 1.92. The most effective compression in history is still deleting half the file.
The result is still impressive: on SWE-bench Verified the Coder scores 77.0 against 81.2 for the original, keeping 94.8%, and 88.57 against 89.71 on LiveCodeBench v6. But experts were picked for code, tool use, long context and vision, and the authors say plainly that degradation outside that set is an accepted cost. For a local coding agent it’s a solid pick. For anything else, take the unpruned build, which weighs 248.8GB at the same IQ2_XXS.
The key number in the experiment where GPT-6 Astra edited Qwen3.5-0.8B’s weights isn’t MATH-500 jumping from 7.9% to 48.1%, it’s answers hitting the 32,768-token limit dropping from 93% to 33%.
An MIT CSAIL team (Takehi, Teoh, Zhang and Khattab) published “Weight-Space Programming” today. Working in a Python kernel with no GPU and no gradient training, Astra traced a direction in MLP layers 15 and 19 that caused reasoning loops and subtracted it from 64 columns in each. That’s 131,072 weights, 0.016% of the model. GSM8K went from 13.8% to 67.1%, IFEval from 21.0% to 62.3%. The model mostly stopped thinking in circles rather than getting smarter, which is still more than most meetings achieve.
The authors say it themselves: an edit can remove an obstacle but can’t add an ability. On Gemma 3 270M, which rarely loops, MATH-500 only moved from 6.8% to 7.7%, and with 10 or 25 practice passes instead of 50 the loops barely budged. If your small local model underperforms, check how often answers hit the token limit before you swap models, because sometimes the problem is a loop, not knowledge.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
OpenAI’s new benchmark for everyday life is missing one thing: any model not made by OpenAI.
LifeVal, announced today, was built from a survey of about 4,000 people, about 100 of whom logged their real tasks in detail. It covers seven categories, from apartment hunting and choosing a kindergarten to government forms and household finances, groceries, meal planning and caring for a sick or aging relative. Instead of one right answer, each task gets its own rubric, like tuition, distance and hours for a school search, or diet, price and availability for groceries. GPT-6 Astra scores 80.6% overall and 100% on tasks like cancelling a subscription by email. Finally an AI that can quit the gym before you ever go.
The set is internal, though, with no public tasks, no leaderboard and no Claude or Gemini scores, so for now it reads more like OpenAI’s progress log than a market comparison. The method is worth stealing anyway. If you’re building an assistant or a customer-facing agent, collect a few dozen real tasks from users and grade each one with its own rubric instead of a single score, because that’s the only way to see where the model actually fails.
For AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them. So we built LifeVal, a new benchmark shaped by feedback from thousands of ChatGPT users on where they most want help from a personal assistant—from finding childcare to accessing government services to managing personal finances.
Over the past year, we’ve used LifeVal to measure and improve how well our models handle these everyday tasks.
openai.com/index/lifeval/
On the older Terminal Bench v2.1, Meituan’s new LongCat-2.5 nearly ties Claude Opus 5 at 88.4 to 89.1, but on the newer v4.0 it trails 38.5 to 53.9.
Meituan shipped the full LongCat-2.5 on November 1: a 1.6T parameter MoE with about 48B active per token, natively multimodal, with a 1M token context. The company says it was trained on domestic accelerators with under 80 GB of memory per chip, and launch pricing is $0.30 per million input tokens and $1.20 for output, with cached input at $0.006. On the old benchmarks everyone is near the ceiling, so the real gap only shows where the tasks are new. Benchmarks age faster than the models training on them.
At that price it’s worth testing, especially for long agentic work, just on your own tasks rather than the launch table. One detail for anyone writing evals: with thinking enabled, temperature is locked at 1.0, and at 0.7 without it, so temperature: 0 won’t give you reproducible runs. Open weights are listed as “coming soon” with no date.
Introducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active.
Built for long-horizon agentic tasks, it brings stronger visual understanding and agentic coding to work ranging from complex coding projects to visual coding such as creative design and 3D tasks.
Learn more about LongCat-2.5: longcat.ai/blog/longcat-2.5/
DeepSeek now lets its agent write plugins for itself, inside a runtime where even the sandbox is a plugin.
Creator mode in DeepSeek Harness, announced today, turns a plain language description into a plugin the agent builds in memory and installs in the same session, from a workspace to a sandbox game to a 3D LLM visualizer. It fits the design, since all of Harness runs on the Cordis framework with an “everything is a plugin” rule: the agent loop, model adapters, tools, sandbox and UI are all swappable plugins. Since its developer preview on August 13, the MIT licensed project has passed 200,000 GitHub stars. A plugin ecosystem that writes its own plugins is the fastest route yet to code nobody has read.
The project’s own docs call it a developer preview with no security audit and breaking API changes between releases. If you try it, test generated plugins in memory before saving them to the plugin manager, and do it in a throwaway directory rather than with write access to your real repo.
Asked at ADIPEC in Abu Dhabi what his biggest challenge is, Sam Altman gave a one-word answer: “Energy.” He said it at the world’s biggest oil and gas conference.
On November 2, the day Abu Dhabi signed four deals with OpenAI, Altman offered a comparison that shows the scale. Six years ago OpenAI’s heaviest internal user consumed about 100,000 tokens a month, and today that’s roughly the average per person on the planet. Energy per query has dropped sharply, but demand climbs faster. Costs are falling in almost every way, he said, “except, you know, atoms are still atoms.” First AI lab CEO who could say “energy” at an oil conference and get applause from both sides of the room.
His most quoted line from the stage, that AI is now “mostly smarter” than our smartest scientists, is a claim without a stated measurement. The energy numbers you can check on a utility bill. If you’re planning AI infrastructure for a company or a country, start with the grid connection and the tariff rather than the model, since you can swap a model in a week and a transformer takes years.
The best number in the “five Qwens beat Fable 5” story isn’t the score, it’s the 23 points the same model gained just from organizing its work.
In the revised GVS5H paper on 100 hard LiveCodeBench problems, Qwen3.8-27B called once solved 63.0%, while the same model orchestrated across roles reached 86.4%. Fable 5 scored 87.4%, so the “win” from the original repo became a statistical tie (p = 0.73). The viral 9.8 vs 8.0 from Memdoor covers 10 problems and 5 passes, far too small to settle anything. Five copies of one model talking through files is basically an open plan office, minus the fights over the AC.
Read the bill too. The revised paper prices a full 100-problem run at $51.75 versus $61.11 for Fable 5, about 15% cheaper rather than the 91% savings people were sharing. Memdoor burned 15.3M output tokens across 50 runs, with a median run taking about 40 minutes. If you run a 27B locally, a simple planner and reviewer split is worth building because the jump is real, just count time and tokens before you cancel the subscription.
💥 Five Qwen3.8-27B agents just beat Fable 5 on LiveCodeBench Hard.
Same model, five roles talking only through files: 9.8 vs 8.0 on 10 hard problems (5 passes).
Tiny sample, so grain of salt. Still a wild local team trick.
via Memdoor 👇
github.com/guregodevo/memdoo…
Qwen3.6-35B-A3B now takes a million tokens of context on a DGX Spark, except the model was trained on 262,144 positions and YaRN stretches the rest.
TensorFold 1.2.1 adds a 1,048,576 token window for this model on CUDA by stretching rope with YaRN at factor 4, plus DFlash2 drafting, which per the release notes hits 130 tok/s on GB10 versus 99 with MTP on a 4.2K prompt. The telling numbers sit in PR #1675: a 300K prompt prefills in 173 s, about 1,740 tok/s, while the full million takes 1,546 s, about 680 tok/s and nearly 26 minutes. Just long enough to make a coffee, drink it and forget what you asked.
So prefill slows roughly 2.5x going from 300K to a million, and the correctness check is a single passphrase placed 5% into the prompt, which proves the window works but says nothing about understanding the whole thing. If you run a local model on a Spark for long documents, treat a million tokens as an overnight batch job, and day to day keep context to a few hundred thousand and reuse the prompt cache.
Within two days, two frontier AI companies admitted the same thing: flat subscriptions don’t survive contact with agents.
On October 31 OpenAI said that from December, ChatGPT and Codex on Plus, Pro and Business will share one token based allowance, with only Instant chat off the meter. A day later Elon Musk wrote plainly that serving Grok Bot was “unsustainably high relative to revenue”, while users on the $300 SuperGrok Heavy plan report burning a week’s allowance in two or three days. That’s an unusually honest admission for this industry, and Musk paired it with a promise of daily fixes instead of silence. The all you can eat buffet works great until the guests bring agents.
An agent doesn’t ask one question. It reads a repo for an hour, calls tools and corrects itself, burning tokens at a scale chat pricing was never built for. If your team runs agents on a subscription, measure tokens per typical task now and keep credits or an API key as plan B, because this isn’t getting cheaper.
Intel just had its best server quarter in 15 years, and its stock lost roughly a quarter of its value in a week.
Data Center and AI revenue jumped 78% year over year to $7.3B in Q3, but PCs fell 5%, and the Q4 outlook ($15B to $16B against roughly $17B expected, gross margin 38% instead of 43%) spooked the market. Lip-Bu Tan called Xeon demand “extraordinary”, so Intel is moving capacity from laptop chips to servers. AI agents are part of that, because every sandbox and tool call runs on a plain CPU. After years of chasing customers, Intel finally has too many of them.
Regular buyers will feel it most. PCs are getting pricier because of the memory shortage, and Intel’s CFO expects weak demand at least through the first half of 2027 with supply tight into 2028. If your company plans a laptop refresh in 2027, compare buying now against stretching current hardware another year, because waiting for cheaper chips may not pay off.
In SemiAnalysis’ new benchmark, not a single model fully solved a finance or legal task, and that says more than the leader’s 12%.
GDPx, published November 2, is 206 agentic tasks across 11 domains, from chip design to quant trading, written by 14 specialist firms. Each model runs in its native harness (Opus 5.5 in Claude Code, Gemini 4 Carbon in Antigravity), gets up to 8 hours per task, and the score averages three attempts. Opus 5.5 leads at 12.0% ahead of Carbon at 10.9%, but under binary grading every model scored zero in finance and legal, so those domains count partial credit only. The lawyers can finish their morning coffee.
The more useful part is what Sundar Pichai pointed at, the cost line. Opus averages $51.41 per task, roughly $4.30 per point of score, while Carbon runs $16.29, about $1.50 per point. If you’re picking a model for long agentic work, price it per unit of finished work rather than per token, and test it in the harness it will actually run in, because GDPx measures the model and its tooling as a pair.
MiniMax is thanking “local builders”, and the smallest MLX build of M3.1 Flash that fits at all needs a Mac with 256GB of memory.
M3.1 Flash is a 428B MoE with about 23B active per token and up to 1M tokens of context. Weights landed on Hugging Face on October 23, and the community quantized them fast: the smallest 2-bit GGUF is 112GB, Unsloth’s is 143.9GB, and the 3-bit MLX build is about 186GB. On a 256GB M5 Ultra that leaves roughly 31GB for context. The people keeping 256GB of RAM at home also heard from Sam Altman this week that it’s wasteful, so the thank-you card arrived right on time.
MiniMax M3.1 Flash, on more of your hardware.
Quantizing a 428B model opens up more ways to run it locally. Lower-bit weights reduce the memory needed to load the model and the weight data moved during decoding. Making those savings useful takes careful calibration, efficient low-bit kernels, and hardware-aware placement of experts across GPU and system RAM.
Thank you to the builders doing that work:
• @UnslothAI @danielhanchen — 2-bit Dynamic GGUF weights at 143.9 GB, with calibration and fidelity testing across coding, tool use and other workloads.
• @ViC305 — SAGE-EXL3, token-level fidelity measurements and a single-DGX-Spark recipe.
• @PeasantSmith — Project Maya quants and expert offloading for local coding and tool use.
• @MiaAI_lab — EXL3-EAQ quantization work.
• @jundotkim + mlx-community — oQ quants and DSpark-powered decoding through oMLX on Apple Silicon.
• @AIatAMD — Quark MXFP4 adapted for AMD Instinct.
• @atomic_chat_hq — GGUFs for local desktop workflows.
• @superalesha — custom inference work combining quantized weights, expert caching and system RAM on 4× RTX 3090.
Every shared quant, kernel, measurement and recipe gives the next builder a better starting point. Thank you.
The alleged Mythos class leak on Hugging Face weighs 7.72 TB, and part of it can be regenerated bit for bit with one line of NumPy.
The burner22/numbat repo went up on October 23 from an account created a week earlier, with a card promising a 14 trillion parameter Anthropic checkpoint and 1,224 MoE experts. On Sunday user FloxUnder showed in the repo discussions that 43 tensors, 28 of them normalization weights, exactly match a random number generator seeded with nothing but each tensor’s name. Norm weights normally start at ones and move during training, so a bit exact match to a seed means those layers were never trained. The most expensive np.random.seed ever, billed in terabytes of bandwidth.
Sure, 43 tensors out of trillions of parameters is a rounding error. But nothing beyond the model card ties this repo to Anthropic, and Anthropic hasn’t said a word. If you pull models from HF for local testing, read the Discussions tab and regenerate one small tensor from its seed before you queue up terabytes, it takes minutes. The Mythos 5.5 sighting in anthropic.com’s code also rests on a screenshot so far, the live page still says 5.1.
43 million Hacker News comments from 20 years boiled down to 6,454 videos, and that ratio is the best part of Andrej Karpathy’s new weekend project.
HackerTube, which he showed on October 31, is YouTube with no algorithm and no accounts: only videos HN commenters recommended since 2007, thumbnails taken from the middle of each video instead of clickbait covers, and a full data export so you can hand it to your own agent. Per his post it took about 2 hours in Claude Code with Opus 5.5 on medium effort. He created a Google Cloud API key and bought the domain himself, and Claude handled the deploy plus hardening the server. The human kept the hardest job, entering a card number.
The split of work matters more than the two hours. Opus pulled and filtered the data, while the cheaper Sonnet got the bulk job of weeding spam out of roughly 40k recommending comments. If you sit on a forum archive, a Slack export or an issue tracker, the pattern transfers as is: the expensive model plans, the cheap one classifies at volume, and the curation is the product.
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org/
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
DeepSeek built an agent designed to take as little human attention as possible, and shipped it with auto accept switched off by default.
It’s dsh-superagent-prototype from the DeepSeek Harness team, open sourced on GitHub on October 31 under MIT. A long running Router watches one repo (and optionally Feishu chats) and hands work to isolated Worker sessions. Every task gets a Ticket with an immutable Mandate, a record of why the agent is allowed to touch it at all. Their starting premise will sound familiar: a human can really follow 3 to 5 sessions, and by the sixth you’re the dispatcher. In their test the Router finished at 15.7 minutes p95 versus 29.5 for a single session with subagents available, which never once called a subagent across 5 runs.
Someone will call it a Claude Code killer. It’s v0.1.0 with one commit, tested on synthetic messages and simulated tools, and nobody measured code quality. But if you run four agent terminals at once, steal the Mandate idea today: one line per task saying who asked and why saves you the morning investigation into where that PR came from.
A 5 minute 34 second Seedance 2.5 film means at least 12 separate generations, because ByteDance’s model tops out at 30 seconds per clip.
That’s the interesting part. Seedance 2.5 shipped in August inside ByteDance’s Jimeng AI and Doubao apps, generates picture and sound in one pass, and takes up to 30 images, 10 video clips and 10 audio files as references. Keeping one character’s fur, face and attitude consistent across a dozen shots comes from that reference stack plus an editor who knew what they were doing. Those eyebrows hold continuity better than most TV shows.
The winners here are small creators who can break a story into shots, since the hard part moved from animating to storyboarding. If you make short films, build a character sheet from a dozen references first and plan every scene in 30 second blocks.