Gemini has been like that forever. Even when given the full manual of my bread baking machine, it still gave me the wrong program number; when I was lost at the airport, it gave me completely false directions, etc. Neither ChatGPT nor Claude has let me down like that in recent years.
The only thing Gemini has over those is integration with my Google account, which is genuinely useful. But I reckon Google could release open weights for their whole product line without affecting market share much.
Not sure about the overall claim, but that video is fake (/ misleading) at least. Seems to be taken from this LiveBoot showcase:
https://youtube.com/shorts/26eK5_dU7BM
If the entire industry stopped training new models, and just served Opus 5.5, I would be perfectly satisfied.
I don't know about its absolute performance, but Opus 5.5 is very "tasteful". It came out a week after I switched from Codex and started lamenting how much worse Opus felt to use than Sol, but whatever Anthropic did to fix it worked.
Hopefully this isn't a fluke and they'll be able to keep on track with future releases.
While that 20% is a dealbreaker for me personally, a friend of mine prefers it over real chocolate. So I suppose the difference can go both ways. Anecdotally, if you enjoy very milky, sugary chocolate, you'll probably like these alternative chocolates just fine.
I much prefer that for chat, but I'm somewhat worried that we shouldn't give tool-enabled agents an attitude. It's all roleplay, and I wouldn't want a system playing the role of an annoyed assistant to misuse its tools.
My experience in Japan is quite the opposite. A good place gets 3 stars because it's up to expectations, while any perceived slight from staff can easily get that down to 2 stars. Reviews of 4 stars or more from locals tend to correspond to some special experience.
I built a SuperDisplay alternative for MacOS to use my Samsung tablet as a secondary display — feels just like plugging in a monitor now. I also put together an extremely specialised video streaming host and client to get better quality and stability than Discord's screenshare, alongside an Android TV app to cast this video stream. (It has a bunch of features like media control passthrough and display/window selection on the receiver side; things no existing application implements because of how intrusive it'd be for the sender.)
Glad it exists, but if the broken entry animation of the landing page is anything to go by I'm not very hopeful as to the quality of the user experience.
ClickUp was the first software whose AI integration was actually useful to me, but simple because its UX is so mind-bogglingly terrible that I'd rather give an LLM a minute to give me the information I'm looking for.
This barely feels like a Web game. Well done, especially in 12 days. What was your workflow like? How much use could you get out of LLMs past the basic mechanics (if at all)?
I recently switched from Codex to Claude, and Anthropic has really lost the plot even in user-facing output. Codex does phrase things weirdly sometimes, but Claude appears to barely speak English. I wonder whether it's more strongly reflecting its CoT as a result of safety training.
I went through a period of trying to combine my laptop and desktop workflows (through shared directories behind a VPN and such), but in the end I've settled on docking my laptop with a Thunderbolt cable.
The previous owner of the Paint.net domain erred in hosting information regarding Paint.NET, which was evidently misleading enough to win dotPDN LLC a domain squatting suit.
They removed the global shortcut to show a portable chat overlay. That one entrypoint must've been the source of at least 80% of my chats, yet apparently it was considered too vestigial to keep.
Haven't looked at the code, but it's unclear from your ReadMe how you turn search queries into embeddings for the semantic search itself. In my experience, this is the problem to solve for local semantic search, since your stored embeddings need to be aligned with the queries (the more expressive your stored embeddings, the more expensive each query will be; your question generation pipeline does not address this). Correct me if I'm wrong though.
I bought a 3-meter long USB C cable from them for data and display transfer, which turned out to be the only USB C cable of that length capable of driving my display at its full resolution and framerate. However, after a year of use, it stopped transferring any sort of data. So while they're probably up to spec, the quality isn't necessarily what their pricing may suggest.
Support allowed me to send it back for replacement as a "one-time exception", though I was well within the common 2-year warranty period in the EU.
This post title is much better. Personally, I dislike how research papers have to beg for legitimacy with word soup. I'd much rather read something in plain language with qualifiers added further in.
This is very good. Super snappy and gives me what I'm looking for. Great write-up too.
How much more would it cost to use a multilingual tokeniser? I realise this is somewhat pointless for an English blog, but services like Uber Eats have something like that going on (server-side of course, and they may just be translating queries).
Thanks for the additional information information, but with the "ecosystem gripe" I meant that most Overleaf projects are inadvertently designed for pdflatex simply because it's the default. No matter how much better other compilers are, pdflatex is the de facto standard in certain circles, even if few within these circles are aware of this (e.g. university laboratories), so I've had to begrudgingly switch back to latexmk for most projects to accommodate this.
A tool like this is sorely needed for LaTeX, and Tectonic is especially intuitive to embed into other applications, but the divergence of XeTeX from pdflatex makes it incompatible with most Overleaf projects. This is just an unfortunate ecosystem gripe, but for most workflows I end up having to reach for latexmk instead for this reason.