tlarkworthy

Oh yeah, I read that article and could not find it agin. Thank you.

It really open my mind to what is special about transformers.

Markov chains learn a fixed distribution, but transformers learn the distribution of distributions and latch onto what the current distribution based on evidence seen so far. So that's where the single shot learning comes from in transformer. Markov chains can't do that, they will not change the underlying distribution as they read.

U put something out on the internet, likely no-one cares, sometimes people will point out a better way of doing it. You gain in that knowledge. I learnt a lot from public critique... it makes you better and more knowledgeable. Harness the crowd, let it out.

well that inspired me to research getting those tests in Germany.

=> heart panel plus

https://en.minu.synlab.ee/heart-panel-plus/

I don't need doctors, I can get ChatGPT to analyse the results.

Yeah I was inspired after https://news.ycombinator.com/item?id=43998472 which is also very concrete

It unblocks that workflow, that's why it's so great. You can have a single script with inline dependencies that are auto installed on execution. That can expand to importing other files, but there is very little setup tax to get started with a script and it does not block expansion.

Interesting but I would like details on how spatial is handled specifically

Cool but 12MB WASM blob. I wish there was a way of making these WASM builds significantly smaller.

These subagents look like tools

the prompt I would like to optimize is the reflection prompt

`You are a prompt‑engineer AI. You will be improving the performance of a prompt by considering recent executions of that prompt against a variate of tasks that were asked by a user. You need to look for ways to improve the SCORE by considering recent executions using that prompt and doing web research on the domain.

Your tasks is to improve the CURRENT PROMPT. You will be given traces of several TASKS using the CURRENT PROMPT and then respond only with the text of the improved using the improve_prompt tool`; const research_msg = `Generate some ideas on how how this prompt might be improved, perhaps using web research\nCURRENT PROMPT:\n${prompt}\n${trace}`

source: https://observablehq.com/@tomlarkworthy/gepa#reflectFn

but I would need quite a few distinct tasks to do that and task setup is the laborious part (getting quicker now I optimized the notebook coding agent).

It's all written up and linked in the notebook and executable in your browser (if you dare to insert your OPEN_AI_KEY, but my results are included assuming you won't).

The evals were coding observable notebook challenges, simple things like create a drop down, but to solve you need to know the observable standard library and some of the unique syntax like "viewof".

There is a table of the cases here https://observablehq.com/@tomlarkworthy/robocoop-eval#cell-2...

So it's important the prompt encodes enough of the programming model. The seed prompt did not, but the reflect function managed to figure it all out. At the top of the notebook is the final optimized prompt which has done a fair bit of research to figure out the programming model using web search.

Recently tried out the new GEPA algorithm for prompt evolution with great results. I think using LLMs to write their own prompt and analyze their trajectories is pretty neat once appropriate guardrails are in place

https://arxiv.org/abs/2507.19457

https://observablehq.com/@tomlarkworthy/gepa

I guess GEPA is still preprint and before this survey but I recommend taking a look due to it's simplicity

They did not ask how wings work. They asked for the bernoulli effect, that's a different question.

We recently migrated from RDS to Iceberg and the savings have been mind-blowing. That said, it's like a deconstructed database and everything you took for granted is now on the outside. Sometimes it's cool, like instead of a Cron inside the database you have a cloud Cron on the outside which is more natively monitor able. Some of it sucks like write contention on inserts so you need a write coordinator lock, and external vacuums. However, the savings are extreme. Better cost per byte stored. Better compression. Less table bloat.

The universal compatibility has not materialised yet. We are still waiting for duckdb and click house predicate pushdown but if that appears we will be able to have specialized query engines over a generic storage layer which would be amazing. Tools like Nimtable underline what compatability unlocks.

I wrote an agent that works in userspace inside the developing program and it frequently reads it's own code to diagnose errors and sometimes tries to upgrade itself, but that causes a hot reload and it loses its own conversation. It does seem to be useful though that it can read it's own tool implementations.

I am making an offline-first thing that serialised to a single file that can be opened without a local Fileserver.

Still working on the UX a little but it seems close to what you want (and I agree). The vision statement is about creating immortal software exactly to fight bitrot https://github.com/tomlarkworthy/lopecode

VR boxing is exhausting. Beat saber at high levels is also pretty sweaty. So yeah, cardio. I also don't play many games on it but it has stuck as a fitness option for years now.

Used mine today for a workout. Super low friction compared to going to a gym.

note three.js [1] has nothing to do with React out of the box though, this page highlights an atypical way of using three.js through a popular React binding.

[1] https://threejs.org/

Yes that's a browser and it implements a standard supported by multiple vendors for multiple operating systems.

no you don't need to run a webserver, it works from a file domain directly. But true its not a CLI. Been thinking about whether that is possible though.

you do not need the internet to open a ".html" file. With some "importmap" gymnastics you can embed images and libraries inside that file, all without the internet, offline-first

https://bsky.app/profile/chrisshank.com/post/3llmhyr33q22x

It was spurs that made me think it was UK centric, lions as well is valid for England too

Llms do not have it memorized. Try asking it to calculate the signed angle between two signed planes. I gave up and told it to copy the unity implementation.

I had this idea too, imaging creating 3d assets using a virtual lathe

No I think you need service workers for notifications. You could do journalling but it's emits a file when saving by default. Index dB works so maybe that's a possability

still a WIP but I will write it up properly soon.

I have made a notebook environment inside a single html file. You have to map dependencies in an importmap to blob URLs and it works. I have a single file website online. And you can click download to play with it locally

https://tomlarkworthy.moldable.app/index.html

You can turn off network and the download works.

I think if you also try to constrain yourself to only building orthogonal components it leads to a needs-based Lego set. I think it is difficult to see the underlying symmetry in a problem space if you design top-down.

oh and now the developer reopened it because it is not actually fully complete, lol. Yep, Iceburg on Clickhouse is WIP. I am actively watching this because it is relevant for my company.