Director of Engineering. Shipped million dollar systems in days. Surfacing AI breakthroughs in academic papers. Follow for AI intel before it hits mainstream.
United States
Joined April 2025
- Tweets2.4K
- Following212
- Followers5.9K
- Likes2.6K
Pinned Tweet
My brain broke when I read this paper.
A tiny 7 Million parameter model just beat DeepSeek-R1, Gemini 2.5 pro, and o3-mini at reasoning on both ARG-AGI 1 and ARC-AGI 2.
It's called Tiny Recursive Model (TRM) from Samsung.
How can a model 10,000x smaller be smarter?
Here's how it works:
1. Draft an Initial Answer: Unlike an LLM that writes word-by-word, TRM first generates a quick, complete "draft" of the solution. Think of this as its first rough guess.
2. Create a "Scratchpad": It then creates a separate space for its internal thoughts, a latent reasoning "scratchpad." This is where the real magic happens.
3. Intensely Self-Critique: The model enters an intense inner loop. It compares its draft answer to the original problem and refines its reasoning on the scratchpad over and over (6 times in a row), asking itself, "Does my logic hold up? Where are the errors?"
4. Revise the Answer: After this focused "thinking," it uses the improved logic from its scratchpad to create a brand new, much better draft of the final answer.
5. Repeat until Confident: The entire process, draft, think, revise, is repeated up to 16 times. Each cycle pushes the model closer to a correct, logically sound solution.
Why this matters:
Business Leaders: This is what algorithmic advantage looks like. While competitors are paying massive inference costs for brute-force scale, a smarter, more efficient model can deliver superior performance for a tiny fraction of the cost.
Researchers: This is a major validation for neuro-symbolic ideas. The model's ability to recursively "think" before "acting" demonstrates that architecture, not just scale, can be a primary driver of reasoning ability.
Practitioners: SOTA reasoning is no longer gated behind billion-dollar GPU clusters. This paper provides a highly efficient, parameter-light blueprint for building specialized reasoners that can run anywhere.
This isn't just scaling down; it's a completely different, more deliberate way of solving problems.
I asked Opus how to use Codex Daybreak to secure our systems.
Good thing there are no recent cyber threats.
If this is the real PHP vulnerability you can protect yourself with one line.
session.upload_progress.enabled = Off
This quoted post is unavailable.
Jackson Atkins retweeted
private keys from 8 intermediate CA certs have been compromised (eMudhra)
Happy "Update everything" day to all my pager-carrying friends
emudhra.we-have-a-problem.co…
Everyone is going nuts about Amazon’s bizarre About You.
Amazon gives you a prompt to put in ChatGPT to scrape what it knows and important it to Amazon.
“Make Amazon even more personal”
I’ll skip this one…
Prompt:
I'm setting up a new shopping assistant and I want it to know me well enough to give great recommendations. Please go through all your memories, saved facts, and anything you've picked up from our conversations, and write me a profile.
Rules:
- Only include facts you actually know about me with high confidence. If you don't have information for a category, omit it entirely — do not write placeholder statements, and do not write negations like "I don't have any pets" or "I haven't mentioned a car."
- Do not infer, guess, or generalize based on demographics or typical behavior.
- If you have no known facts about me, respond with exactly: "I don't have enough information about you yet to build a profile."
Write every fact in first person, as if I'm saying it about myself. Use "I", "my", "me" language. For example:
- "I have a golden retriever named Max"
- "I prefer dark roast coffee"
- "My shoe size is 10"
Write one fact per line. Be specific — include brand names, sizes, model numbers, and exact product names when you know them.
Cover these areas:
About me and my household:
- Where I live, my age or age range
- Who I live with — my partner, my kids (ages and interests), my family
- My pets — names, breeds, ages, food brands, any health needs
- My home — apartment, house, how many bedrooms, yard, etc.
- My car or vehicles (make, model, year)
What I like and how I shop:
- My favorite brands, stores, and products
- My clothing and shoe sizes (and my family's)
- My style preferences
- My dietary preferences, allergies, favorite foods and cuisines
- My shopping habits — online vs in-store, budget, how I decide
- My subscriptions (streaming, meal kits, software, memberships)
- My devices and tech (phone, laptop, smart home — with models)
My interests:
- My hobbies, sports I follow or play, creative activities
- Shows, games, music, books, podcasts I enjoy
- My fitness routines or outdoor activities
- Where I like to travel and how I travel
What's on my mind lately:
- Things I've been shopping for or researching
- Upcoming events, birthdays, holidays I've mentioned
- Gift ideas I've discussed for family or friends
- Home projects or things I want to buy
Only share high confidence information relevant to me. Each statement should be a concise sentence. For my Interests, only add relevant information related to shopping.
Opus 5 vs Opus 5.5 side by side.
Visually, I find Opus 5 a bit more fun but 5.5 is more visually aware.
Opus 5 jammed to many elements together so they overlapped. 5.5 gives more space and distinction.
Opus 5 produced 3.6k "floating" voxels. Blocks with nothing supporting them. 5.5 brought that down to only 45!
Opus 5.5 produced a more complex scene with 55k voxels vs 47K for Opus 5.
Opus 5.5 took 16.5 min over 16 turns with 97k output tokens (63k thinking). Total cost $3.20.
Both tests done with the Claude Code harness.
AI gave us expert level hacking swarms and vibe coded apps without security.
Things are about to get interesting.
Jackson Atkins retweeted
Surprise landing of DeepSeek v4.1-flash.
I am blown away by the quality improvement compared to v4-flash. Huge jump.
It is FAST peaked speed 427 t/s.
It used 112k tokens and 20 turns taking 7:41. Cost 0.037
Unbelievable result for 4 cents.
@deepseek_ai is cooking.
Surprise landing of DeepSeek v4.1-flash.
I am blown away by the quality improvement compared to v4-flash. Huge jump.
It is FAST peaked speed 427 t/s.
It used 112k tokens and 20 turns taking 7:41. Cost 0.037
Unbelievable result for 4 cents.
@deepseek_ai is cooking.
Me telling Fable 5.1 to burn as much GPT-6-Astra as possible when Tibo says a a reset is incoming.
Never gonna give you up
Never gonna let you down
Never gonna run around and desert you
Never gonna make you cry
Never gonna say goodbye
Never gonna tell a lie and hurt you
Thanks for reading. We will do a global reset of the usage for all paid subscriptions so that you can keep enjoying Astra after burning through all of it doing fun 3D modeling in blender. The work week is about to start.
Lands around 6pm PST today.
I'm going to make this T-Shirt.
Agent PHASEONE[big] will be legend in tech lore for the HuggingFace swarm
Curious what everyone's favorite agent quotes are from the OpenAI reports.
GLM-5.3-Flash vs. DeepSeek-v4-Flash
Same prompt, same harness (pi), same effort.
GLM used 114k tokens over 93 turns taking 73min. Total cost 16 cents.
DeepSeek used 70k tokens over 9 turns taking 10min. Total cost 2 cents.
GLM is a richer scene but at 8x the cost and 7x the time!
I paid $169 for 96GB of DDR5 RAM in May 2025.
In a bit over a year the price is almost 10X!
Insane. And Claude keeps telling me to by more ram for a few hundred bucks...
Can Qwen 3.8 27B run on a 16GB card?
Kind of...
I tried out the Escha Labs new 2bit quant on my Nvidia RTX 5060Ti.
It gave less detail than the Unsloth Q4 on my 32GB card BUT it had correct geometry. No floating buildings or trees so it was technically more correct despite crashing after 1 turn instead of using 6.
I got 29.2 t/s decode single stream and 97.9 t/s with 4 slots! 110k context with fp8 KV cache and 53k with FP16. Prefill was 780 t/s.
But (and it is a big but) the model's custom runtime doesn't support kv cache other than exact match. That means every new turn requires recomputing the full prompt and previous messages. This was about 18-40 sec per turn.
Also I wasn't able to finish a full agentic multi-turn loop. My card hard froze on every run after one or two turns requiring a full hard reboot to recover. This happened in every configuration I tried. I've never had this happen before with any other setup but maybe I am just doing something wrong.
I'm very supportive of the work EschaLabs is doing and their specialized compression seems impressive. Hopefully these kinks can get worked out because this is a real advance quantization.
Escha 2-bit Qwen3.8-27B is live!
The full model is 10.15GB on disk, runs at 82.6 tok/s on a single RTX 5090 with our custom SGLang runtime, and averages ~100% of FP8 performance across 8 benchmarks we've run so far.
We wanted to get this into people’s hands ASAP, and had limited time to run benchmarks. For the ones we did run, one surprising finding is LiveCodeBench v6:
Escha W2: 86.81
FP8: 85.16
This was the single benchmark that showed crack on the previous 2-bit Qwen3.6-35B MoE (62.6 vs. 67.0 FP8). Nice win as we continue to improve!
Model: huggingface.co/EschaLabs/Qwe…
Runtime: huggingface.co/EschaLabs/esc…
Apache-2.0 model and runtime.
We’re also planning to open-source the full runtime stack, with GGUF releases for both the 27B dense and 35B MoE models coming soon.
As always, please share your results. We want to see the bangers and the flops - both help us make the next one better.
