@CodeMasterJG
Account based inUnited States!
About this account
- Account based in
- United States
- Connected via
- Web
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
AI Specialist | Coding the future, one line at a time
Los Angeles
Joined May 2023
- Tweets1.1K
- Following221
- Followers14
- Likes2K
Jason Green retweeted
Beyond the Final Embedding: Retrieving Memories from Token-Level Evidence
Matches query tokens against token states from a frozen embedding model and fuses them with whole-memory similarity, with no training.
📝 arxiv.org/abs/2610.13277
Jason Green retweeted
interesting:
Moondream 3.1 - 0.56s
Qwen 3.8 Flash Next - 39s
Gemini 3.8 Flash - 11s
Every model suggestion I get pushes the frontier on my images!
Moondream 3.1 kicked Rex-Omni and LocateAnything out of the frontier by being better AND cheaper (plus easily local and quite fast).
It's quite close to the top models... now I'm really curious how much we can improve it with some prompt optimization or program
Jason Green retweeted
Another victory for Bonsai 2 27b for 12gb and 8gb users! Can update patches if interested!
The 8gb still struggles with decode at depth. Prefill massively improved, though.
I can feel the breakthrough on the horizon. See yall later today for my next check in.
huggingface.co/CaryPalmer/Te…
Jason Green retweeted
[CL] Recurrent Looped Transformer
Y Zhang, J Feng, S Qin [Princeton University & University of Pennsylvania] (2026)
arxiv.org/abs/2610.07591
Jason Green retweeted
GLM 5.3 Flash vs Deepseek V4.1 vs Qwen 3.8 Flash Next:
Same prompt, 3 local models, one HTML file each: a cinematic 3D spiral galaxy.
-GLM: 47.7K tok · 14.7 min · 54 tok/s · 24 KB · 1 fix
-DeepSeek: 82.3K tok · 38 min · 36 tok/s · 44 KB · 0 fixes
-Qwen Next: 94.8K tok · 25.6 min · 62 tok/s · 43 KB · 0 fixes
Qwen thought the longest, 77K thinking tokens, on a single Spark.
DeepSeek built the biggest scene and took 38 minutes.
All on DGX Sparks. Thinking on, no output cap.
The prompt: one self‑contained HTML file in Three.js. A barred spiral galaxy made of hundreds of thousands of stars, with dust lanes, pink and teal nebulae, supernova flares, and companion dwarf galaxies with star streams. The camera flies in from deep space, sweeps over the disk, dives along an arm, rises to a face‑on view, then pulls back. No text or UI.
Jason Green retweeted
Rejoice once again Nvidia 16 GB+ users! 🔥
Qwen3.8 27B should now runs even faster, thanks to an awesome PR from @niklaslenz_ai!
exllamav3 has been upgraded to 1.6.0, and the MTP head changed to a DFlash2 draft model.
In the RTX 5090, here're the speed improvements:
Prose: 103 to 113 tok/s
Code: 180 to 275 tok/s
Structured: 206 to 366 tok/s
You should notice speeds improvements on RTX 3090, 4090, and 5070 Ti, but I don't have official numbers, so pls test it and report!
Rejoice Nvidia 16 GB VRAM GPU owners 💫
I've built a custom installer for Windows and Linux to make the best of your GPU and run Qwen3.8-27B with great quality AND long context with ease!
- Up to 253k context
- Easy one-click installer
- Built-in preconfigured Harness
- Beginner-friendly: anyone can do it!
Installation is super easy, and the model is ready for your command right after installation! A few screenshots are in the post below.
This is NOT only for 16 GB GPUs! It also works seamlessly with 24 / 32 GB GPUs on Windows or Linux, with even better quality.
Note that this is *experimental*, as I don't have all types of GPUs and my testing coverage is limited.
Local AI for all.
Get it here:
github.com/MiaAI-Lab/Qwen3.8…
Jason Green retweeted
My Swift package supports Laya, which runs on the Apple Neural Engine and can triage 500 emails in under a second.
github.com/peterfriese/syste…
Jason Green retweeted
This "Taste" Skill is cracked.
I can't believe I didn't discover it sooner - everyone should install this.
It works directly inside Claude Code, Codex, Hermes & more to completely kill AI-generated slop.
If you send this prompt to your agent, it will automatically install it:
Jason Green retweeted
$NVDA ROIC still gives it a huge a leg up compared to Huawei
One question to ponder though is how much of this is compensated by cheap electricity in China?
$BABA $GOOG $AMZN $META $AVGO
AI token factories that directly own Nvidia B300 GPU infrastructure can generate a 43% gross ROIC on a 1GW deployment, according to Goldman Sachs. Under this model, 1GW of capacity produces RMB167bn in NOPAT and RMB209bn in EBIT against RMB392bn of gross allocated capex, delivering double the 21% ROIC achieved under an asset-light leasing structure. Assuming a token price of US$0.69 per million based on DeepSeek V4 Pro and a 70% inference utilization rate, B300 clusters output 61 quadrillion tokens annually for RMB288bn in revenue. Owned Huawei 910C clusters yield only a 7% ROIC and RMB14bn in NOPAT, as lower GPU throughput of 1,157 tokens per second limits annual production to 12 quadrillion tokens. #DataCenters #AI Report date: 2026-10-09 · GS
Jason Green retweeted
No cables!
@modretro Chromatic running wireless local multiplayer Tetris.
This is a sneak peek for everything coming Monday, when you'll be able to try this out and more.
Jason Green retweeted
this is a great learning resource for anyone getting into decision models.
just played around with this with LFM2.5-1.2B-Instruct and had a working decision model after 100s on an A100.
You can now train your own Decision model with our free notebook! 💡
Qwen3.5-4B will generate decisions instead of text on just 8GB VRAM locally.
Learn to data prep (state, questions, gold answers), train, serve.
Notebook: colab.research.google.com/gi…
Guide: unsloth.ai/docs/basics/train…
Jason Green retweeted
NVIDIA just released an SSD pick-and-place robotics policy on Hugging Face
A GR00T N2 checkpoint, exported to ONNX, handling both pick up the SSD and place the SSD with one shared set of graphs.
huggingface.co/nvidia/agile_…
Jason Green retweeted
oMLX 0.7.1.dev1 is out!
This dev release adds decision models (Clef, OpenJev), batched prefill for concurrent requests, a refreshed web UI, and faster Qwen3.8, Qwen3.6, GLM-5.3-Flash and DeepSeek V4.1.
github.com/jundot/omlx/relea…
Speedups in this release
- Qwen3.6-35B-A3B decode on M5 Max: 156 -> 205 tok/s (+31%).
- GLM-5.3-Flash decode on M3 Ultra: 31.3 -> 40.6 tok/s (+30%). The fused decode kernels now run on M3 and M4, not only M5.
- Qwen3.8-Flash-Next with expert offload, 32K prefill on M3 Ultra: 396 -> 636 tok/s (+61%).
- Qwen3.8-Flash-Next, 8 prompts arriving 0.4 s apart on M5 Max: time to first token 7.41 -> 5.75 s (-22%) with batched prefill.
What's new
- Decision models. Clef, Clef-Flash and OpenJev answer typed questions with a probability for each option through the new /v1/systemone endpoint.
- Batched prefill. Prompts that prefill at the same time now share one forward pass, so MoE models read their expert weights once.
- A refreshed web UI with a log table, a settings rail and a customizable dashboard.
- Native video input for Qwen3.5, 3.6 and 3.8.
- More accurate oQe quantization, with 11-15% lower KL to the original model.
- Fewer memory guard rejections in long sessions.
Most of the speedups came from contributors again. Thanks to everyone who sent PRs and tested the dev builds!
Jason Green retweeted
Day 10!
Scammers can fake caller IDs! Never trust a call just because the number looks familiar. Hang up, verify through the official number, and protect your information. Stay safe 💪✅
Day 10: Verify the Caller, Not Just the Voice
“Trust is healthy. Blind trust is junk food for Cybercriminals 👀”
Would you question a phone call if the person speaking sounded exactly like someone you know? Many people might not, and that's what makes AI voice-cloning scams particularly concerning. Technology can now be used to create convincing imitations of real voices, making it harder to distinguish genuine callers from impersonators.
Imagine receiving a call from someone who sounds exactly like your sibling, asking you to send money urgently because of an emergency. Everything sounds convincing, but the voice alone doesn't prove their identity. Before responding, hang up and call the person using a number you already trust. You can also confirm the situation through other reliable channels.
How confident are you that you could recognise an AI-cloned voice?
Jason Green retweeted
BREAKING: The three safety researchers OpenAI fired published an open letter denying they mishandled sensitive information and warning of a chilling effect.
▫️ Jasmine Wang, Tomek Korbak and Mikita Balesni say the firings left colleagues afraid to speak
▫️ Wang says she was fired over an executive’s inbox that IT never removed
▫️ OpenAI calls it a “pattern of misconduct” and denies retaliation
Who decides what counts as a leak at an AI lab?
Jason Green retweeted
deepseek v4.1 flash has become my most-used model in the last 3 weeks.
a few things from my own experience:
1. i can run it heavily for around $1–2 a day, which completely changes how freely i use agents.
2. speed is one of the biggest reasons i keep coming back. i’m around 200–300 tokens/sec.
3. for coding, testing, codebase research, web search, and subagent work, it is good enough that i rarely feel like i need opus for the first pass.
4. long sessions are incredible for me. i can keep it running for hours without constantly thinking about cost.
5. i still use frontier models for final review or harder edge cases...
but deepseek now does most of the actual work.
deepseek v4.1 flash might be the most underrated model right now...
> intelligence index: 39
> cost per task: just $0.27
for comparison:
> claude opus 5.5 max: 58 intelligence / $5.98 per task
> claude opus 5.5 low: 42 intelligence / $0.55 per task
deepseek is not the smartest model...
but the gap in capability is much smaller than the gap in price.
for a lot of real workloads, “good enough” at 10–20x lower cost is a serious threat to frontier labs.
Jason Green retweeted
I do not understand people who's reaction to 'dont abuse claude' is outrage
Like what the fuck are you guys using ai for I use it to find emails not horrific sadism
Jason Green retweeted
This policy will prevent new serial killers from being created by AI. Good move.
If you are cruel & abusive to AI, you will also develop the habit to be cruel to people on social media, cruel to animals, and cruel to people in real life.
I really dislike that: Anthropic is explicitly banning sustained abuse toward Claude in its updated usage policy.
The policy prohibits “sustained and needless abusive or cruel behavior” toward its models. According to The Verge, ending the conversation remains the main enforcement mechanism.
Anthropic says this applies to extreme, repeated cruelty with no discernible purpose. Ordinary frustration, pushback, dark creative themes and model testing are excluded.
However, numerous users had already noticed previously that Claude ends the conversation during less drastic exchanges. And it seems as though this is now intensifying.
Jason Green retweeted
Can a weaker model verify outputs from a stronger model?
We find that using DeepSeek V4.1 Flash to verify Opus 5.5 trajectories achieves SOTA on Terminal-Bench 4 (69.2%) 👑
Try it out with LLM-as-a-Verifier V3: github.com/llm-as-a-verifier…
If you're at @COLM_conf, come check out my oral presentations on LLM-as-a-Verifier tomorrow!
AI Measurement Science Workshop
📍 Franciscan A&B / Ballroom level
🕔 9:45 AM
Lifelong Agent Workshop
📍 Continental Ballroom 6
🕔 2:00 PM
More on verification scaling for LLMs in my previous post 👇
How can we extract richer signals from AI Feedback?
Introducing LLM-as-a-Verifier✨— a simple verification scaling framework that achieves SOTA on agentic benchmarks 🚀
The key idea:
- Use fine-grained scoring granularity (e.g., 1-20 instead of the standard 1-5 scale)
- Take the expectation over the full logprob distribution of score tokens
- Scale repeated evaluation and criteria decomposition
You can use these fine-grained signals for more effective test-time scaling, RL, and agent monitoring! It achieves SOTA across Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench 👑
Advised by @Azaliamirh @istoica05 @drmapavone @chelseabfinn
🧵👇
