mmargenot

I'm a little confused by your distinction between vendor and lab. AI2 (Allen Institute for AI) is a western lab and they did release a fully open source model, with open weights, data, and training code (Olmo3).

Their technical report on the release is excellent: https://allenai.org/blog/olmo3

They're a public research institute as opposed to one of the proprietary frontier labs, but they've been very open about the science.

Many Qwen model weights are licensed with Apache 2.0 (though unfortunately not the latest releases). Unfortunately open-weight is generally the term being used for weights-available, so the battle tends to be on whether the models are "open source", i.e. whether they release training code and data alongside the weights.

You’re trying to get calibrated confidence, otherwise it’s less effective than you’d hope.

This approach with more data for calibration is basically the beginning of making a decision model like Jev.

Very fun to see moltbook but trying to be actually productive.

Are there any agent-sourced or agent-proposed problems or is it primarily the moderation staff that directs the pillars of agent “attention”?

The existing state is definitely couched in using robust seed data and spinning it into magnitudes more tokens. I'm not sure what a world that isn't 100% grounded in objective human-written truth looks like, but I suspect the next step is generating synthetic data from synthetic data that has been blessed by human beings, and moving yet further away from that with more sampling and evaluation.

There are some things that are true (I hope), so using those concepts, approximated by a human text filter, as seed seems reasonable. Hopefully the diversity in sampling methodologies and opinions about what makes a meaty token pays some dividends down the line for model diversity.

LLMs can be pre-trained on LLM outputs, and there have been a few recent interesting implementations of this.

- https://arxiv.org/abs/2609.37891

- https://arxiv.org/pdf/2508.10975

- https://www.datologyai.com/blog/eating-the-rainbow-a-primer-...

As well as recent model releases that used large fractions of synthetic data (Arcee Trinity):

- https://arxiv.org/pdf/2602.17004v1

It's mostly a matter of increasingly complex and sophisticated sampling methodologies to sift for the valuable synthetic tokens.

There have also been some interesting results about models that use the same data converging to very similar behavior, but I can't find the specific reference at the moment. That one tracks intuitively with the notion that the model is building a latent representation of everything it sees.

Did something change with it for depth of integration? Goose has been able to use local models via OpenAI compatible endpoints for at least a year

This was great! When you think about optimizing prompts with GEPA (or comparable methods and tools), do you consider each tool or skill separately? How do you think about the optimization of the system prompt for a large agentic system?

I imagine that you do a collection of passes to cover each overlapping set of what you want evaluated, but the system prompt makes all cases dependent on each other.

What I’ve done in the past is use the system prompt to extract subjective criteria for an LLM judge (like various system prompt statements that contribute to brand voice) and check individual traces with that for evaluation, but I’d like to move beyond including that in a prompt at all.

What I prefer is to be able to iterate and iterate on my spec, then implement once. That’s what I’ve come to as a process with CC, and I’ve gotten really good results. Whenever I spend less than 20-30 minutes actively discussing the spec with the agent, the quality of implementation rapidly drops, even with some tiny things.

I like the overall idea of this, but the writing of the spec to me is more like building the thing than the implementation itself, and needs more active human (and agent) cycles.

basically, yeah. wild flash in the pan, but i think it's cute in that it's signaling reduced down to the absolute bare bones.

not a lot of horror assets either unfortunately. i wanted ghosts and ghouls and zombie tiki with my pirates but got industrial wasteland and candyland.

A key part of maintaining a second brain is revisiting and pruning it. Much like a todo list, without that practice it absolutely just accumulates bunk.

In recent news this page that's been somewhat viral recently seems an interesting successor to the million dollar homepage: https://outbid.lol/

A little nihilistic, but I find it charming.

And you get bm25 for free with so many modern setups! I do still love to experiment with tuning semantic search for your specific corpus via various kinds of embeddings, but bm25 is hard to beat.

This is so cool! For the time being I’m still locked into notion for my handwritten knowledge base, but I love this for incorporating external information.

For the semantic search is there any chunking/processing that happens with the content or do you need to be diligent about having a large embedding context (and/or small content)?

as a fun aside here from the abstract it does come down to how you use it:

> AI users who maintain similar homework completion time as non-AI users experience small learning losses.

If the pure completion time of homework goes drops significantly with an LLM tool, I suspect these students are spending the extra time turning the material over in different ways to internalize, which is interesting.

Glad to have scientific results on this, though in my view you could get this from first principles. Homework is onerous but it forces you to learn the material and get it into a configuration that works in your head, which you then validate with the exam.

I think my notetaking undergoes a transformation each year. Last summer I made a remembrance agent plugin in obsidian which has been useful for collecting my notes on technical info and cross referencing. I still have a notion zettelkasten, but it’s become more of a tool for my creative writing. I expect this year to take the leap on building things more like remembrance agents to enhance my recall when doing creative work.

I think people get tied to particular tools and methods and optimizing them just right, rather than adapting things piecemeal in a way that works for them.

Sounds like good support for a few different paradigms.

Python has pretty good autogenerated reference docs, but if you want more than markdownified docstrings it makes them as difficult to maintain as the other pillars. As a set, the pillars aren’t as neatly grouped in Python documentation tooling, but I typically differentiate how-to, tutorial, etc with type tags in Notion or other indexes that point to where each type of thing lives.

I love whenever this comes up. It’s a great framework for thinking about and writing docs.

It’s hard to keep documentation up to date, however, and I find that items like tutorials and reference materials (unless generated off versioned code) can drift pretty far over time.

A feature I like in concept that notion introduced with wikis ages ago was a “verification” timestamp, where you specify a timeframe after which the doc owner has to reconfirm that the doc is up to date. A bit too easy to rubber stamp, unfortunately, but maybe if it took the doc offline entirely without a more robust audit.

I just retried a few queries that were failing ~3 days ago and they are returning correct [enough] results on mobile as of today, when I don't need them. I'm not sure of the cause or how to actually trigger a reindex or a fix when I need it.

Gmail's search on mobile, where things come up in the quick results only to disappear when I complete the search, is a millstone about my neck.

I find this routing problem to be opaque and I’m generally skeptical that the label people are trying to predict is meaningful.

If you really need more discrimination of the complexity of an input to get an efficient response, sft or rl tuning something for your harness would be more effective.

How do you handle the more densely written pages in script ? I did a very similar exercise OCRing works from this exact collection, but I stuck with the English books for the first pass.

I love this approach, very Diamond Age. I uploaded a picture of my cat and learned a lot about coat genetics. I think it fixated on his coloration.

AI applications that would help normal people in a significant way are pretty lacking, so I'm not surprised. So much conversation about AI products is cycles of "this tech will change everything" without material backup outside of coding agents.

Here’s an explicit example of the above from today using the above dataset: https://x.com/kaiostephens/status/2040396678176362540?s=46

This is a great idea! I saw a similar (inverse) idea the other day for pooling compute (https://github.com/michaelneale/mesh-llm). What are you doing for compute in the backend? Are you locked into a cohort from month to month?

It is more common now to improve models in agentic systems "in the loop" with reinforcement learning. Anthropic is [very likely] doing this in the backend to systematically improve the performance of their models specifically with their tools. I've done this with Goose at Block with more classic post-training approaches because it was before RL really hit the mainstream as an approach for this.

If you want to look at some of the tooling and process for this, check out verifiers (https://github.com/PrimeIntellect-ai/verifiers), hermes (https://github.com/nousresearch/hermes-agent) and accompanying trace datasets (https://huggingface.co/datasets/kai-os/carnice-glm5-hermes-t...), and other open source tools and harnesses.