In light of the recent controversies around Meta’s AI bot of xAI’s Ani, but also the broader debate on model sycophancy, I’m reposting this

Anthropic’s recent evals for the release of Sonnet 4 and Opus 4 represent a good opportunity to re-share a blogpost from last year entitled “Artificial Intelligence : what everyone can agree on”

It touched on benchmark-gaming

I’ve been testing the new Google AI Studio feature “Build” as a promising mitigant to Gemini’s code slop tendency (which the AI Studio UI makes a horrible experience)

Unfortunately this is very far from ready to do actual work

Strongly advise to wait a couple of weeks (details below)

If I believe the marketing materials, the latest Claude models can work continuously for longer than the Twitter backend does

Earlier this month, I wrote about the Cambrian explosion in the voice modality (TTS and STS) and the evolution of frontier labs’ commercial strategies

www.pitti.io/blogs/not-to...

One of my mutuals here suggested a Bluesky Wednesday as a way to revive the interest for this platform.

Reposting here a blog series bringing together various thoughts and fragments posted elsewhere around AI alignment.

Some context first

www.pitti.io/blogs/aligne...

I’ve not found this platform very welcoming last time but I’m willing to give it another try.

Reposting here an essay from early Feb on existential economic questions for generations X, Y and Z.

Touching on:
- Aging population
- Sovereign debt
- Technology
- A new world order

www.pitti.io/articles/exi...

This place is clearly not a safe haven as its decentralized nature is a double-edged sword

AI seems even more divisive than on other platforms. No punches pulled over here (and it’s not great)

In September, I wrote about what everyone can agree on.

www.pitti.io/articles/ai-...

Qwen2.5-Coder-32B doesn’t really live up to expectations

Can’t handle a redux store in a fairly simple app with 3 components

When a new model comes out and I want to vibecheck it with a single prompt, my prompt is :

Please provide a detailed analysis of the strengths and weaknesses of the business model of a typical hairdresser

The screenshot comes from my FAMLX repo where I explored the entropix adaptive sampler. I listed some tokens that appeared important to me in each tokenizer

Here is a presentation of the project (not focused on entropix)
www.pitti.io/projects/famlx

I was yesterday at a symposium organized by the Institut Presaje (which I actively support) at the Cour d’appel de Paris for a discussion on deployed AI.

What’s particularly interesting about this sector is that they approach AI from the bottom (users) and the top (legal/regulatory oversight)

I can see why people want to scrape bluesky data at scale but I doubt that it will be of significant value beyond RL to imitate the style short messages

But I do believe that an open API is of significant value, at the very least to extract my own content and do whatever I want with it.

Gemini is not a very helpful assistant here

Major grammatical error, I’m not sure I’ve ever seem something like this

One of the byproducts of my attempt to find appropriate settings for the entropix sampler with different MLX models is that I ended up replicating parts of the “softmax is not enough” paper across several models.

Qwen2.5 0.5B is very prone to that, much more than the larger Qwen2.5 models

It’s not so often that you have an opportunity to start from a clean slate. I’m excited about this one

Quick intro about PITTI and the person behind the posts (I/me): PITTI is a personal project that I started several years ago to post about my side projects and areas of interest outside finance