bogtog

In fact I think long-term autonomy (in the range of several hours) and self-correcting is going to be where we see most improvements in coming years.

Right, model intelligence defines the scope of things they can one shot

I also suspect that users naturally calibrate to a model's useful scope, gradually getting positive/negative feedback and gradually making their requests bigger/smaller than before

I'm surprised I never heard people talking about using -Pro variants, even though their rates ($125-175/M?) aren't drastically larger than old Opus ($75/M), which people seemed to use

I'd be curious if there were some measurements of the final effects, since presumably models wont <think> in caveman speak nor code like that

Not listed here is how banks themselves have changed to be almost entirely online

Sorry what? Was this not the central theme of the article? (albeit with a title that used the word "iPhone" to be catchier)

This is wrong. It's not insider trading. Lutnick didn't have inside information. His son just had a brain. Anyone who read the case knew which way the court was going, it was the least surprising decision ever. Perhaps the only surprising thing is that the court ever heard it.

If this was so obvious, wouldn't there have been more competitors pushing down the value of it?

Mr. Less-than-Consistently-Candid strikes again

But now that most code is written by LLMs, it's as "hard" for the LLM to write Python as it is to write Rust/Go

The LLM still benefits from the abstraction provided by Python (fewer tokens and less cognitive load). I could see a pipeline working where one model writes in Python or so, then another model is tasked to compile it into a more performant language

I figure OP would try and give the models pure text forms of the game?

.....

l....

l....

l.ttt

l..t.

This is fair, but this seems like the only way to test this type of thing while avoiding the risk of harassing tons of farmers with AI emails. In the end, the performance will be judged on how much of a human harness is given

I associate "yello" with Homer Simpson: https://www.facebook.com/TheDoctorZaius/videos/7233283715092...

(fingers crossed I'm not somehow doxxing myself by sharing a fb link)

People will pay extra for Opus over Sonnet and often describe the $200 Max plan as cheap because of the time it saves. Paying for a somewhat better harness follows the same logic

The game looks really good, although I think it'd be improved if the sphere was a bit smaller. It feels like it takes too long for the game to become difficult

Oh my reasoning was coming at this from a different angle: H200s were released in November of 2023, so they're over 2 years old at this point while still being valuable

A few months ago, there was a lot of news lambasting tech companies for extending the depreciation lifespan of GPUs from ~3 years to ~5 years. Do these price hikes suggest a longer lifespan is probably the right way to see how long these GPUs will be valuable?

Thanks for sharing. I'm surprised you can't just ctrl-a + copy-paste your bank statement and get it to work easily

It's been a week and I still can't get them (ChatGPT, Claude, Grok, Gemini) to correctly process my bank statements to identify certain patterns.

Can you give any more details on what you mean? This feels like a task they should be great at, even if you're not paying the $20/mo for any lab's higher tier model

I don't think the commentor above is saying that an AI should necessarily apply the redaction. Rather, an AI can serve as an objective-ish way of determining what should be redacted. This seems somewhat analogous to how (non-AI) models can we used to evaluate how gerrymandered a map is

That's fair. I sometimes find myself pausing or just talking in circles as I'm deciding what I want. I think when I'm speaking, I feel freer to use less precise/formal descriptions, but the model can still correctly interpret the technical meaning

In either case, different strokes for different folks, and what ultimately matters is whether you get good results. I think the upside is high, so I broadly suggest people try it out

Claude on macOS and iOS have native voice to text transcription

Yeah, Claude/ChatGPT/Gemini all offer this, although Gemini's is basically unusable because it will immediately send the message if you stop talking for a few seconds

I imagine you totally could use the app transcript and paste it in, but keeping the friction to an absolute minimum (e.g., just needing to press one hotkey) feels nice

I'm using Wispr flow, but I've also tried Superwhisper. Both are fine. I have a convenient hotkey to start/end recording with one hand. Having it just need one hand is nice. I'm using this with the Claude Code vscode extension in Cursor. If you go down this route, the Claude Code instance should be moved into a separate window outside your main editor or else it'll flicker a lot

There are a few apps nowadays for voice transcription. I've used Wispr Flow and Superwhisper, and both seem good. You can map some hotkey (e.g., ctrl + windows) to start recording, then when you press it again to stop, it'll get pasted into whatever text box you have open

Superwhisper offers some AI post-processing of the text (e.g., making nice bullets or grammar), but this doesn't seem necessary and just makes things a bit slower

Using voice transcription is nice for fully expressing what you want, so the model doesn't need to make guesses. I'm often voicing 500-word prompts. If you talk in a winding way that looks awkward when in text, that's fine. The model will almost certainly be able to tell what you mean. Using voice-to-text is my biggest suggestion for people who want to use AI for programming

(I'm not a particularly slow typer. I can go 70-90 WPM on a typing test. However, this speed drops quickly when I need to also think about what I'm saying. Typing that fast is also kinda tiring, whereas talking/thinking at 100-120 WPM feels comfortable. In general, I think just this lowered friction makes me much more willing to fully describe what I want)

You can also ask it, "do you have any questions?" I find that saying "if you have any questions, ask me, otherwise go ahead and build this" rarely produces questions for me. However, if I say "Make a plan and ask me any questions you may have" then it usually has a few questions

I've also found a lot of success when I tell Claude Code to emulate on some specific piece of code I've previously written, either within the same project or something I've pasted in

It tearing when I waved my mouse around was a nice surprise

Opening that video, American-style pickup trucks are about 40% more likely to kill a pedestrian 100% more likely to kill a child (the video argues that this mostly stems from the shape of the front). These cars also get into more crashes

Honestly, banning these things seems sensible when the only thing going for them for most buyers is seemingly an appreciation of their style

The premise of this post and the one cited near the start (https://www.tobyord.com/writing/inefficiency-of-reinforcemen...) is that RL involves just 1 bit of learning for a rollout, rewarding success/failure.

However, the way I'm seeing this is that a RL rollout may involve, say, 100 small decisions out of a pool of 1,000 possible decisions. Each training step, will slightly upregulate/downregulate a given training step in the step's condition. There will be uncertainty about which decision was helpful/harmful -- we only have 1 bit of information after all -- but this setup where many steps are slowly learned across many examples seems like it would lend itself well to generalization (e.g., instead of 1 bit in one context, you get a hundred 0.01 bit insights across 100 contexts). There may be some benefits not captured by comparing the number of bits relative to pretraining.

As the blog says, "Fewer bits, sure, but very valuable bits", this also seems like a different factor that would also be true. Learning these small decisions may be vastly more valuable for producing accurate outputs than learning through pretraining.

There aren't many major labs, and they each claim to want AI to benefit humanity. They cannot entirely control how others use their APIs, but I would like their mainline chatbots to not be overly sycophantic and generally to not try and foster human-AI friendships. I can't imagine any realistic legislation, but it would be nice if the few labs just did this on their own accord (or were at least shamed more for not doing so)

For 5.1-thinking, they show that 90th-percentile-length conversations are have 71% longer reasoning and 10th-percentile-length ones are 57% shorter

Unfortunately, I also don't want other people to interact with a sycophantic robot friend, yet my picker only applies to my conversation

GPT-OSS-20B at 4- or 8-bits is probably your best bet? Qwen3-30b-a3b probably the next best option. Maybe there exists some 1.7 or 2 bit version of GPT-OSS-120B

They report benchmarks on the huggingface page (https://huggingface.co/utter-project/EuroLLM-9B)

They almost exclusively compare their model to prior models from 2024 or older and brag about "results comparable to Gemma-2-9B". I'm not sure what I expected. The eurollm.io homepage states "EuroLLM outperforms similar-sized models", which just seems like a lie for all practical purposes

An overly charitable interpretation is that EuroLLM isn't a reasoning model and has minimal post-training, so they sought out comparisons to such models (they're still ignoring reasoning models that have non-reasoning modes)