armcat

Great writeup! I've done some work on Codecarbon [1] in the past, it's a pretty standard tool for measuring ML/AI workload emissions, it would be awesome to have your findings contributed to that project as well.

[1] https://github.com/mlco2/codecarbon

Thank you!

I’ve been doing this with voice mode in agent harnesses. In Codex for example I often speak orally and I find that (1) I discover something new myself, (2) agents seem to understand much better than pure text prompts. The only downside is it costs lot of tokens.

From my time building ML models (2016-2023) evaluation and interpretation was a large chunk of it. I often used "brute force" interpretation methods but there were others like SHAP [1] and LIME [2]. The other large chunk of ML work focused on feature engineering and data prep. Today in the LLM agentics era this is out the window. There are "evals" but no one does it seriously, and interpretation is often propagated to the consumer.

[1] https://shap.readthedocs.io/en/latest/

[2] https://christophm.github.io/interpretable-ml-book/lime.html

This is fantastic. I feel the KV cache system is basically a must for any serious inference hosting, both for the provider as well as the consumer. I guess they have this but just haven’t open sourced it yet.

Does anyone know if this is based on the NIST report on the "Evaluation of DeepSeek AI Models" [1]?

> DeepSeek models advance Chinese Communist Party (CCP) narratives much more frequently than U.S. models.

[1] https://www.nist.gov/system/files/documents/2025/09/30/CAISI...

I get "Sorry, this content is not available in your region." when trying to access this link. EU/Sweden.

Gibson's quote "the future is already here - it's just not evenly distributed" continues to hold true, especially today. Amazing author and unparalleled mind's eye.

I feel the hardware vibecoding is the next and currently untapped frontier. I’ve noticed AI labs are placing more emphasis on hardware as well.

Reinforces what Karpathy tweeted [1], that AI is effectively tackling 2-5% of work and the gap between those fully exploiting AI and barely touching it is growing. The question remains whether this AI focus on the small % of workers is actually intentional (by tech/labs), or a natural limit of where AI is right now.

[1] https://x.com/karpathy/status/2109361546505966046?s=20

The toaster is next!

"Reflection’s first model, Beam, scores 30 on the Artificial Analysis Intelligence Index, placing it #2 among US open weights models" [1]

[1] https://x.com/ArtificialAnlys/status/2111087606759030831

Looks like a potentially solid coding model, but this is also interesting:

"Reaches comparable solution quality to leading open models using approximately 3x fewer inference-time reasoning tokens."

This is very cool, I can see a big use case on this for streaming analytics and anomaly detection. So if I get this right, you get:

- Amortized O(1) work per window update

- O(1) aggregate queries, assuming a fixed cost merge operation

- O(W) memory for W buckets, with sketch size treated as fixed

- No need to recompute the entire window every second

For Codex there is no need to decompile because unlike Claude Code, Codex is open source [1] and written in Rust.

There are lot of other amazing open source agent harnesses like OpenCode [2] and Pi [3].

[1] https://github.com/openai/codex

[2] https://github.com/anomalyco/opencode

[3] https://github.com/earendil-works/pi

There is no actual Claude Code source code in there, if you actually look in the repo.

What should be stated here clearly is that these are open weight models [1] with MIT license. The models range from 5M to 200M parameters, and are best-in-class for each model size, and are best-in-class overall based on the public benchmarks.

[1] https://huggingface.co/collections/mistralai/timestral

They've done considerable effort in making sure this model can run on basically every model runner, from Unsloth to MLX, https://dev.meta.ai/docs/muse-spark-open/deploy

Absolutely, it is part of the critical infrastructure now, since everything is in the cloud, including certain software from other critical infrastructure (transport, energy, banking). In the Iran conflict both Oracle and AWS infrastructure was targeted.

Is this a US phenomenon? In Sweden dental care is free up to 21 years of age, and for adults heavily subsidised. I heard similar stories from my friends in US about every aspect of the “healthcare system”, from pregnancy and birthing to seeing someone when you have a pain in your back. Seems like the whole thing is incentivised to suck you dry

The best thing about this article is that it's now inspired me to deploy some small LLMs across all my appliances and turn my house into a distributed AI cloud.

Two concerns here: (1) That "rogue AI" is going to be a catch all and an excuse for all poor security and due diligence; (2) vibecoding in itself, particularly by ppl who have no idea how the software works and are entirely focused only on the outcomes, is going to create lots of security holes.

And then there is the "real" rogue AI agents. And it's difficult now to see what is what.

There are a tonne of AI companion apps out there now, lot of them focused on nsfw, it's going to be interesting to see if character.ai is an isolated case, or if it's a domino effect.

This is interesting from the perspective of scale (because Claude is pumping out massive volume of code). I am still lacking semantic re-invention of source control for agentic era. As others have pointed out - what is the point of human source control in the era where it's just agents building things. We may need to rethink this completely - in terms of "intent" (what was the agent trying to ACTUALLY accomplish), "provenance" (which agent made what decision - based on what evidence), and "continuous verification" (can correctness be maintained through ongoing validation, repair, and rollback rather than a single pre-merge testing gate).

These are just some of the things - there may be many more.

This is very cool and on a lower level (more in the trenches) compared to what I did earlier in the year (i did a small writeup and built some skills [1]).

Because I am so-and-so pixel artist, the biggest unlock for me has been in rapidly iterating with graphics assets and sprites, and it's actually helped me become a better artist and animator (I built my own little tool for this [2]). The rest Codex or Claude Code seem to handle incredibly well (plus I have 30 years of programming behind me).

[1] https://github.com/acatovic/gothicvania-codex-demo

[2] https://github.com/acatovic/ai-game-studio

I applaud you on the experiment, this is a great showcase in how the latent space is shared between modalities and a great performance achieved.

You stated WER of 5% - against which benchmark?

Whistle [1] for example is a 16MB STT model but WER varies between 4-26% depending on the benchmark.

[1] https://cactuscompute.com/blog/whistle

Karpathy yesterday posted [1] about the gap between small % of population getting high leverage from AI, and the rest of the world. We have a much more urgent gap between software powering the "physical world", and the digital natives. The digital natives now have 10x leverage with AI, with the ability to rapidly find and deploy exploits. We're now seeing a lot of this - fridges defrosting, access cards for construction workers not working. Next is train and air traffic controls, bank transactions/clearing wiped out, shipping being dismantled and so on. The gap is now being fully exposed.

[1] https://x.com/karpathy/status/2109361546505966046?s=20

One feedback - for the integer values like number of storeys and bays, it would probably be better with like a "+/-" or a slider. But just my own UX opinion ;-)

They use the OCR + structured data extraction in this example, but I'm thinking you could extrapolate this to any scenario where you are extracting some data (from a table, markdown file, OCRed doc), and then use a decision model to perform a fast on-the-fly review/validation.

This is beautiful and as an aside, I love the clean minimalist UI as well!