clacla👹dev

Ideas, code, shipping. Building and whatever catches my attention next. AI, crypto, bootstrapping.

[clacla👹dev @clacladev]

All right, I am ready for Claude Opus 5

[clacla👹dev @clacladev]

I really like this kind of experimentation. Programming languages will be soon an agent tool, not a human tool anymore. So why bother with syntax. Just the concepts represented as an AST. Maybe even lower than that

[Chris Tate @ctatedev]

In the next version of Zerolang

Agents get closer to the compiler

Instead of making agents edit source text, then recover meaning through format/check/build/test loops

Zerolang makes the compiler's semantic graph the program

Agents query it, patch it, and get checked edits

[clacla👹dev @clacladev]

PumpFun is regularly creating the cheap poisonous fast food like experiences that crypto is rightfully criticised for. First memecoins then this… let’s create a better future for all participants not just for selected few

[Crypto Sum @cryptosum_]

Pumpfun launches bounties and tasks for users to create and wage crypto on

[clacla👹dev @clacladev]

@mattia_asti I am going to give it a go

[clacla👹dev @clacladev]

In this new agentic workflows era, is there a kanban-like tool where for each column the tickets get processed by a skill and then pushed forward?
E.g. Research -> PRD -> Issues -> Dev -> QA -> Deploy

[clacla👹dev @clacladev]

I've implemented GPT-2 on my machine. Some takeaways:
1. wow
2. impressive that a few simple concepts, in the right shape, can be this powerful
3. I still dislike python and its tooling... 🥶

[clacla👹dev @clacladev]

Today was the first day in a loooong time I was able to dedicate 6h to study. I mean like reading something, not learning by doing. I am gassed out

[clacla👹dev @clacladev]

I am studying token embeddings for LLMs and how we represent meaning of a word and the context of the sentence it's in. It's mind boggling to learn that we can pack so much meaning in a series of numbers, enriched, refined and summarised over multiple steps.

[clacla👹dev @clacladev]

I will dedicate next week to go way deeper in how frontier LLMs and AI systems work, get architected and trained.

I bought this course https://t.co/2zPOZb0z9z and brewed a big jug of coffee. I am ready

[clacla👹dev @clacladev]

I finally got around recording a demo of Juno the AI Defender app I am working on.

[clacla👹dev @clacladev]

@ajwarner90 @rudolf6_ Agree. The price being down for so long in one reason per se. It attracts less people.

And I confirm that the fact that the only working use cases are around trading makes it a bit boring…

I want to see more fun and creative use cases

[clacla👹dev @clacladev]

Ethereum continues shipping and making progress to be the internet and finance backbone of the future. Not that bs chain like Solana. Relentless focus on fundamentals and not yapping, sooner or later will pay off in its price.

[amanda.eth @amandacassatt]

Everyone should be concerned about Ethereum.

[clacla👹dev @clacladev]

I just released a little tool to understand if your website is getting indexed by ChatGPT, Google AI, Perplexity, etc, or if you have to update your settings. Let me know if you would like to see more data integrated
https://t.co/4PuuJcyavY

[clacla👹dev @clacladev]

I am thinking of joining this Technical AI Safety course https://t.co/UyXXcysBjU
Any feedback?

[clacla👹dev @clacladev]

I just discovered that when my MacBook M1 has the neural engine chip drawing max current it makes a noise at 6khz (buzz).

Every time a local ai model is running, the chip is maxed out, and my Mac buzzes... 💀

[clacla👹dev @clacladev]

Evaluation of different models for webpage threat detection completed.

gemma3:4b => good quality and fast (~2.5s)
gemma4:e2b => better quality, but too slow (~8s)
qwen2.5:3b => low quality
llama3.2:3b => lowest quality

We have a winner for a baseline machine (M1 Max).

[clacla👹dev @clacladev]

Eval suite progress

[clacla👹dev @clacladev]

And I am adding a debug time feature to add new cases to the eval corpus. So I can simply click a button to add the current email/message/webpage to the eval tests corpus from Juno UI (while fixing the threat levels).

It could be post-training dataset in the future

[clacla👹dev @clacladev]

I am creating a full Swift evals suite to Juno AI defender.

I need to understand which [input preprocessing + ai model + system prompt] leads to better threat detection.

I am measuring latencies, cost of mistakes (false positives are worse than false negatives), prompt inj

[clacla👹dev @clacladev]

Yesterday I added the ability to specify which apps you want to monitor with Juno (local ai defender). There are strong defaults ofc, but maybe I want to remove noise...