dr_kiszonka

Maybe because it is not such an easy problem to solve? With so many people working of quantizing and fine tuning open models, I don't buy that it is a 4-8 hour task.

I would also be genuinely greatful if someone could post how they fixed the looping issue through sampling settings without significantly sacrificing reasoning effort and context size.

Fixing this problem would result in more tickets from accounts flagged as "bad actors." So first you pay for fixing the problem, and then you pay to resolve the tickets. Maybe nobody at LinkedIn wants to pay for it?

It sucks for those banned erroneously.

Your experience reminded me of Terminal Bench 4, where Terra 5.6 is much closer to Luna 6 then Sol 6. https://benchlm.ai/benchmarks/terminal-bench-4.

Looking at newer models, Terra 6.1 is only a hair behind Sol 6.1, hopefully making it a useful Sol replacement for slightly easier tasks.

Reviewing 150k loc line-by-line is not practical. I think we need a new set of tools to facilitate such tasks.

Otherwise, we will be spending most of our time monitoring AI's code writing process.

I am curious about Pi, but my experience with a seemingly more fully-featured Hermes has been endless tinkering and fixing minor issues. Have your experience been at all similar?

I don't know if people hate it, but I don't use it because it is overpriced (you need the $50 pro subscription to do more than tinker) and it had multiple CVEs this year. There are also a lot of n8n "influencers" on YouTube, but that is not n8n's fault.

I think it is fine for prototyping if someone prefers a GUI.

Is "the whole thing" referring to the "current system"or something else? Asking out of curiosity, as I am sadly not a younger American.

Nice. I will give it a try too. Thanks!

Handy worked fine for me on Windows, but on Fedora, I can't get it to work at all.

Unless it is something with the playback on my end, your animations and scrolling are a bit jerky too.

I like this. It means a human was involved.

We can get refurbished phones with 8 GB LPDDR5 for less than regular DDR5 modules.

LPDDR5 is slower but maybe a small cluster could be useful?

There isn't much choice, I am afraid. It is Chinese models or Gemma or Llama? (I am skipping a few lesser known ones.)

I take this opportunity to shame GitHub for completely ignoring the mobile experience in their own Android app. The project's README is pages upon pages of largely blank space. In general, the app does not render mermaid diagrams and does not allow for zooming in, so smaller pictures are unusable. Even if you access pictures directly in a repo's source, you still can't zoom in. iPython Notebooks, which are extremely common in data science, are an "unsupported file type" and are not rendered.

(OP, nice project! Sorry for my rant.)

Thanks for sharing this! These vendor lock-in attempts are very annoying.

Unsloth documented the vision components degrading more during quantization for one of the earlier Qwen models.

Looks good but consider referencing CHOMPI clearly and sufficiently prominently on your page, not just in the repo.

Contributors: GCSS had the original idea. JPP tried to talk him out of it. JPP did the first literature search but GCSS lost it. GCSS drafted the manuscript but JPP deleted all the best jokes. GCSS is the guarantor, and JPP says it serves him right.

But the chat version does change over time, correct? It has been my experience that Sol's performance has deteriorated significantly.

Very curious that everyone here (so far) seems to assume this dashboard presents real data.

It looks fun! Reminds me a bit of TLDRAW computer.

Same here. 5.5 would sometimes oneshot relatively complex models even without MCP. Opus, however, even with Fusion 360 MCP access would usually make comically bad models. It seemed that Opus tried to reason through the problem while GPT would come up with a general plan of using whatever tools were available to solve the problem. In chat, it would use the available Python and trimesh. It was also the only time I had ChatGPT work on something for over 30 min without giving up.

Thanks!

Thank you for your work on Snap! and for finding the time to respond to our comments.

What device did you get your nephew? When I was looking for one, those listed on the Make Code website were pretty bad and expensive (or unavailable). I would love to find an affordable retro handheld like Miyoo or Anbernic with Make Code support.

This is much better than I anticipated. I don't mind AI that much, but this version of HN surfaces really interesting articles I've missed.

So I know a Lisp dialect? Fascinating.

The vocals in the few songs I listened to are so very clearly AI. I find Suno's much more credible in many genres.

If anyone has tried Suno's Studio 2, I would love to hear your take on it. At $24-30/month, it's approaching Ableton Live 12 Suite rent-to-own territory, without the "own" part.

Could anyone explain to me the difference between thinking traces ("intermediate tokens") and the final responses? Specifically, why is it that Claude Opus 5's reasoning in Code is very easy to follow and sounds quite natural, while its answers are full of these very annoying AI-isms and sentence fragments that are void of meaning?

Are thinking traces and final answers trained for different objectives?

I spotted Claude doing it when accessing PubMed.

We also have sanctions and incentives to induce specific behaviors, but they don't apply to agents. We can put a muzzle on Guile 3.8 but we can't turn it into Genuine 3..8.