- > Opus 4.5 is absolutely a state of the art model.
> See: https://artificialanalysis.ai
The field moves fast. Per artificialanalysis, Opus 4.5 is currently behind GPT-5.2 (x-high) and Gemini 3 Pro. Even Google's cheaper Gemini 3 Flash model seems to be slightly ahead of Opus 4.5.
- > Safari isn't cross platform either
WebKit is[1][2].
[1]: https://webkit.org/downloads/ [2]: https://webkit.org/webkit-on-windows/
- It's just a long winded way of saying "tied embeddings"[1]. IIRC, GPT-2, BERT, Gemma 2, Gemma 3, some of the smaller Qwen models and many more architectures use weight tied input/output embeddings.
- Incidentally, in Adlerian psychology; all problems are considered people problems.
- > BFL likely should have held off the release until their Apache 2.0 distilled model was released in order to better differentiate from Nano Banana/Nano Banana Pro.
Qwen-Image-Edit-2511 is going to be released next week. And it will be Apache 2.0 licensed. I suspect that was one of the factors in the decision to release FLUX.2 this week.
- An AI agent isn't an evolutionary optimization loop.
- That's very reminiscent of the idea behind the SAM (Sharpness Aware Minimization) family of optimizers.
- > This company is far from "open source", it's had over $1B USD in funding.
Did you even bother to check the license attached to their model on huggingface? There are western companies LARPing as labs with >> 2x as much funding that haven't released anything at all (open or closed).
- > that’s what a coding agent is too.
Not quite.
- Please read section 2 of the paper[1] cited in the blog post. LLMs are used as a mutation function in an evolutionary loop. LLMs are certainly an enabler, but IMO, evolutionary optimization is what deserves credit in this case.
- > What the author calls criminal is the way hospitals typically bill Medicare and private insurance providers.
Interestingly enough, the FBI considers double billing and phantom billing by medical providers, to be fraud.
https://www.fbi.gov/investigate/white-collar-crime/health-ca...
- Ternary quantized weights for LLMs are a thing. Most of the weights in Bitnet b1.58[1] class models[2][3] are ternary (-1/0/1).
[1]: https://arxiv.org/abs/2402.17764
- Controlled fusion has always been 30 years away.
- Perhaps we're overestimating human intelligence and underestimating animal intelligence. Also funny that current LLMs are incapable of continual learning themselves.
- I wear it for about 60-90 minutes and then take it off for a 10-15 minute break. 2-3 such sessions, every day. Also, I use environments all the time.
- I've been using it almost every day for Mac Virtual display, for about 1.5 years now. I don't have any other use case for it.
- As with almost everything else in CS, it's a tradeoff. Pre-fill is compute bound, decoding is memory bandwidth bound. Speculative decoding works when the draft model is more often right that wrong, because most architectures have a lot more compute, compared to memory bandwidth.
- As are systolic arrays, see Google's TPUs.
- Biscuits and crisps are generally considered to be ultra processed foods, despite containing only basic ingredients.
- It used to be a native app.
- 1Password used to be good 10 years ago, but not anymore. A couple of days ago, there was a post about Electron based apps that slow down macOS Tahoe (due to older versions of Electron using an undocumented API). When I ran the script on my laptop, 1Password was on the top of the list.
> 1Password.app: Electron 37.3.1 (Contents/Frameworks/Electron Framework.framework/Versions/A/Electron Framework)
[1]: https://www.hackerneue.com/item?id=45437112
Edit: Judging by the downvotes, it looks like there are a lot of electron lovers here. Why the hate for more efficient native apps? Are bloated binaries, janky UI and lower battery life, features? :)
- A fish rots from the head down.
- I’ve been on both sides of the table. To me, all of those are far more palatable than petty company politics (both in BigTech and startups).
- There's a way out. Build your own company and make it something beneath you.
- Yeah, I only posted two links from my notes, from when I was looking at this a few months ago. Here's one on MIG.
- Link to their huggingface account with (some of) the model weights. I couldn't find a link to it on their website, white paper or GitHub.
- The cycle time seems rather long, in the order of months to years (See table 1 on pg 18 of the paper).
Who is working on AI agents startup to automate this? /s
- There's a community maintained fork.
[1]: https://en.wikipedia.org/wiki/Inflection_AI