petercooper

I don't really have a definitive answer, so this is more thinking out loud, but why do the Economist, Spectator, WSJ, NYT, or the like continue to do reasonably well?

News is everywhere now, and AI can do a reasonable job guiding me on any economic or political topic I might care about. Yet.. named authors with track records or strong personalities/opinions still seem to draw readers and have value, and maybe CSS Tricks is mildly like that in some niches.

I say this with a slight tongue in cheek, but Mistral does have a better option: they recently added GLM 5.3 to their platform (and it works rather well). Not trained in the EU, though, of course ;-)

As someone who was intrigued by these new machines but also has no interest in running Windows, I'm quite happy to learn it works. Though admittedly I was expecting to see DGX OS first.

To be fair, Deno and Bun's enhancements seemed to do a lot to set a fire under Node and have it implement various things it didn't have when both projects launched. This is a good thing.

A bit frivolous, but I once used a spare VPS like that as a one-listener Internet radio station for my mom which worked with her portable radio (she has no interest in using Spotify etc.)

Smaller models have been able to do these sorts of tasks, but a little slower, for a while now. Give a small Qwen 3.8 model a classification task and force a structured output, and it'll do a good job. I've used Qwen 0.8b for basic image classification in <500ms on my local machine for a while now.

There are a few technical details that can reduce the latency significantly (covered in the post) but the real insight has been from watching the reaction to Jev and seeing that there's enough of a market interest to offer it as a distinct thing. The underlying concept/approach was already there.

You can also go beyond Jev. Qwen 3.5 0.8B is fantastic at basic image classification/question answering (including OCR elements) also. Though rather than looking at logits, I get it to output a structured JSON object and it does simple object classification tasks on a Mac at under 500ms a pop (I forget how far, but I think it's like ~250ms) with good accuracy (depending on task).

I've had similar thoughts! I grew up around a heavy smoker. My dad told me to never smoke and I never have. But I randomly tried a nicotine lozenge in my 40s and found it such a "lightbulb on" moment that I've been using pouches for a few years now and my executive functioning still feels better for it (though not as much now as in the first two years, alas).

I had a similar thing with Xcode the other day. I unticked the "download a code assistance model" thing, clicked to continue and it started downloading it anyway (which, to be fair, you can just X out of).

There might be some practical applications of this sort of idea like in situations where you want/have a heavily restricted vocabulary to build from. You can already do this with LLMs but they can get very... "distressed" if you force logits, whereas this would not.

Come to think of it, I'm now curious if it would do well at building SQL queries, say.

I believe the compensation part only kicks in if it’s the airline’s fault.

CLaude, in particular, quietly, but genuinely and deliberately, overuses adverbs.

FWIW, on my Mac Studio I get ~24-27 tok/s generation between 0-16k context in - that's on the Q6_K GGUF with speculative decoding on. I have spent zero effort optimizing/improving this so far but will be trying the 4 bit MLX next (I've tended to find models drop off somewhat below 6 bit but maybe that isn't the case nowadays).

How's the compute side now, I wonder? Because while the Ultras have impressive memory bandwidth for inference, processing prompts still takes a dog's age on my M3 Ultra. I heard the M5 makes some strides forward in this area, though, and the M7 in particular promises to go a lot further.

More likely double that, even. I think you'd still see many buyers there. You can spend like $16k alone on a RTX 6000 PRO with a mere 96GB of VRAM now..

They seem to be suffering from the supply constraints like everyone else. They phased out the higher capacities on the M3 Ultra Mac Studio a while ago, and if you order a 128GB MBP, say, you're looking at six weeks or more for delivery.

Allegedly, they aren't making a big deal out of the M6 which won't have Pro, Max or Ultra configs at all, instead waiting for the M7: https://9to5mac.com/2026/08/08/apple-m7-chip-heres-why-it-ma...

It reminds me conceptually of the idea of using a ST:TNG replicator to just give you another replicator of your own, or asking a stereotypical genie for "infinite wishes". The genie is indeed out of the bottle in many ways.

As well as the headline in/out changes, people heavily using agentic coding tools will want to note the 6x (off peak) and 12x (peak) increase to cache hit pricing on Pro (since cache hit can easily make up 90%+ of input on long sessions).

DeepSeek was hugely underpricing cache hit pricing before and even after this increase they're still cheaper on that metric than every other provider I'm aware of, but it will put an end to those "I used 1 billion tokens and spent $4" reports.

And confirmed: https://x.com/deepseek_ai/status/2087864589895798968

Pro's cache hit prices are 12x or 6x higher now depending on time of day.

"quietly" is the biggest AI writing tell of Q3 2026.

Not to take away from the broader point, but modafinil is lumped in with methylphenidate as a stimulant that "increase[s] the activity of neurotransmitters such as dopamine and norepinephrine". Modafinil does also stimulate the orexin system and is part of why it's so effective for wakefulness and not as a stimulant in the medium/long-term.

No, it’s a licensed black cab thing. London has had mini cabs as well for many decades who didn’t need to study anything and were not even licensed till about 20 years ago. However, their downside is they have to be prebooked and can not be hailed on the street. Apps count as booking and not hailing I guess.

I agree with you to an extent, but you have certainly given me food for thought.

Sticking to LLMs, they seemingly get their intelligence (whatever that really means) from building models rich with knowledge, so you could have a point. But Qwen models seem to be particularly good, even at small model sizes, at maintaining both their own knowledge while acquiescing to and integrating external information in the moment.

Hopefully this boils down to the smaller versions they've teased. In my experience, Qwen models are the closest to the "less knowledge, more intelligence" (yes, the two are hugely correlated!) ideal some tool-dependent tasks need. Even the 3.5 2B can be easily prompted to always lean on tools and not jump to false conclusions (although its actual coding skills are abysmal, as you'd expect).

I'm guessing this is almost entirely about the incredibly low cache read prices. Few have come close to them, nothing has a bigger effect on (a typical) session price, and with the price of RAM right now, they have to be the biggest pain point for them right now? A 10x increase in cache read would be a significant increase, yet would still keep them cheaper than every other provider of their model (at least based on the prices at https://openrouter.ai/deepseek/deepseek-v4-pro#providers)

Shopping with my kids I noticed we kinda have a replacement here in the UK in the shape of HMV (originally a spin off of His Master's Voice). When I was a kid it was mostly a record store, but now it's focused on what I'd call geek/"fandom" tat: posters, figurines, comics, band merch, etc. Whenever I'm dragged into one, it's always full of geek/fandom types hanging out.