swiftcoder

So the question becomes, how many other parts of the inference pipeline have left 1000x optimization opportunities lying on the table?

I've had the impression that last few years that it was basically a solid field of tech bros at these things

This is art

Neither. They ran all the cases using both models, picked the winning result for each one, and then said "if you had a router that guessed with 100% accuracy, here's what it would have picked"

To be fair, even the most efficient hybrid f150 only gets up to 23mpg. The raptor is the worst, but none of the models are great.

They are increasing in Europe, yes, and unlike the American southwest, any sort of artificial cooling is pretty uncommon in most of Europe.

On the flip side, the cost of cooling is steadily falling due to government incentives to replace oil-fired boilers with heat pumps, and the rapid growth of solar driving down daytime market prices for electricity.

Notably, no one has gone to jail yet. Figure its only a matter of time till someone in the AI world gets caught defrauding investors though...

Not sure how that translates at all into "stuff" being more affordable...

I'm pretty sure "stuff" in this context is mostly manufactured goods, and those have fallen in absolute price. Examples abound in the tech sector: even adjusted for inflation, a MacBook Neo is the cheapest Mac ever released, and so on...

Hah. Yeah. Weirdly, most software managers also seem to believe this whole "senior writes a plan, junior implements" is cost-effective (it typically isn't, one should invest in hiring/promoting more seniors).

Those people can carry around a cheap LED lamp if it so pleases them...

The frontier labs have been tightly controlling model access lately to prevent further distillation, and the Chinese labs are still catching up rapidly - it's likely that whatever moat they used distillation to overcome early on is already dry

Particularly if Anthropic or OpenAI are throwing credits at you, which isn't that uncommon for US startups at the moment

Some do, some don't. As a concrete example, the F150 [1] pretty reliably doesn't. Big SUVs like the Jeep Wrangler can be pretty far down there too.

[1] https://www.fueleconomy.gov/feg/noframes/50104.shtml

That’s a weird thing to include. That sounds like an argument _for_ more aggressive lights. I didn’t expect to see a “think of the surveillance system”

One of the biggest excuses for ultra-bright streetlights is crime prevention (usually shortly followed by CCTV). I'm not really sure why we continue to indulge that point of view, but if one is indulging it...

I suggest you not lookup the difference in price between an iPhone in the EU vs the US, then...

You get 12mpg?

It's the USA, so he's probably dropping the kids to school in a pickup truck...

Curious whether they will upstream windows support. WebKit really needs a decent rendering backend on Windows, would unlock a whole world

You want a unique ID so that you can track which units are experiencing issues, you want private keys to be able to securely connect to your backend service and/or mobile companion app.

A local-only device with no telemetry can probably skip some of those

Yeah, it's important that your own people are doing regular inspections of the production line, keys get burned at the final station (use temp keys for QA stages before that point), and you retain software control of the HSM.

You probably aren't going to do this whole rigmarole for small runs of devices like in the OP.

Depends on where the ICs originally come from. No point shipping your ICs from China -> US, flashing them, shipping them back US -> China for assembly, and then ship them China -> US again to sell

Almost all commercial devices need a unique ID, private key, etc burned in on the production line. For a big enough production run, you'd likely ship an HSM to the factory

It’s not that the Chinese firms are any less likely to misuse your data, it’s that you don’t live in china, so their abuse of your data is unlikely to directly impact your day-to-day life in the same way

Some orgs also use them to extract concessions from dependencies further up the chain. Upstream service won't fix an issue that's been causing problems for us? "unintentionally" let it become a SEV, so that we can send the CoE up the chain and get Jassy to drop the hammer on that team...

I think poultry are excluded by the title ("land mammals"). If they were there, it appears[1] they should be about equivalent to the total biomass of sheep.

[1] https://www.pnas.org/doi/full/10.1073/pnas.1711842115

assuming focus is a limited resource

Focus (on a society-wide level) is not so very limited. One of the strengths of our whole society being fragmented into a bunch of distinct organisations, is that they can all go off and solve different problems at the same time.

Shouldn’t problems be addressed in order of importance?

Most of the really big problems facing society can only be addressed through multi-solving (i.e. something like climate change isn't going to just hang around waiting for us to solve one sub-problem at a time). Not to mention, the solution space is often extremely interconnected - for example, finding alternatives so that society can eat less meat helps address both climate change and the horrors of factory farming at the same time.

Feel like a lot of comments are missing the point here - the authors weren't trying to make a great music video, they were trying to test independent tool use.

Despite how awful the end result may be, the models did successfully plan and execute a moderately complicated generation and editing process.

Counterpoint: this is the same instructions provided for a wide class of printed optical illusions, likely well represented in the training set.

It's interesting how zoom levels affect legibility. The first example is so large on my main monitor that I can only read the decoy text. Zooming out reveals the actual text. Which implies that all one would have to do is teach the LLM to downsample it once or twice, and it would then be able to read it...

It's not reasonable to expect the application layer to carefully partition its work into "I/O heavy" and "CPU heavy" parts.

If your application has both of these and doesn't partition them, you already have a fundamental flaw in your application architecture. Every serious piece of software does this, from web GUIs, to servers, all the way to video games.

It's not reasonable to queue up an arbitrary amount of work without back-pressure

Absolutely, which is why we use pull-futures instead of push-futures, across the entire rust ecosystem. Creating unbounded queues in Tokio requires a fair bit of error on the part of the programmer

If you tell it to write a spec -> then write the tests -> then implement, the LLM should be able to pretty much one-shot a compiler frontend. LLMs really benefit from the kind of task that has a built-in validation loop.