In his last blog post, Sam Altman also revealed how much power the average chatgpt query uses, and it's in the same ballpark.
> People are often curious about how much energy a ChatGPT query uses; the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.
Interesting point about PC going digital-only as Nintendo is a fascinating counter-example.
While they offer digital downloads on the eShop, their pricing actively discourages it.
Case in point: I just bought my kid a new first-party Switch game. Physical copy on Amazon was ~25% cheaper than the identical digital version on Nintendo's own eShop. Even my 9-year-old noted how illogical it seems, the physical version requires manufacturing, shipping, retail markup, yet costs significantly less than the digital bits that have near-zero marginal cost.
It strongly suggests Nintendo wants the physical retail channel to thrive, or values the perceived permanence/resale value of cartridges.
This context makes the Switch 2 "gamekey" cartridges (physical auth token, digital download) fit their pattern of valuing a physical artifact and retail presence, even if the data delivery shifts.
If I understand it correctly, that's a valid concern but the way structured generation library like outlines[1] work is that they can generate multiple variants of the inference (which they call beam search).
One beam could be "This is a way to solv-". With no obvious "good" next token.
Another beam could be "This way is solv-". With "ing" as the obvious next token.
I can’t say for certain, but I’d guess that writing without the letter “e” is slightly more difficult in French than in English. For one, “e” is a bit more common in French (around 15% of all letters, versus about 12% in English). But more importantly, French grammar adds extra challenges—like gender agreement, where feminine forms often require an “e”, and the frequent use of articles like le and les, which become unusable.
That said, I think the most impressive achievement is the English translation of the French novel. Writing an original constrained novel is hard enough, but translating one means you can’t just steer the story wherever you like. You have to preserve the plot, tone, and themes of the original, all while respecting a completely different set of linguistic limitations. That’s a remarkable balancing act.
What I will add is that constrained generation is supported by the major inference engine like llama.cpp, vllm and the likes, so what you are describing is actually trivial on locally hosted models, you just have to provide a regex that prevent them to use the letter 'e' in the output.
OSS model do not have to be local models, and it's not just about privacy, imo.
DeepSeek R1 hosting is out of reach for most, but it being open is a game changer if you are a building a business that needs the SoTA capabilities of such a large model, not because you will necessarily host it yourself, but because you can't be locked out of using it.
If you build your business on top of OpenAI, and they decide they don't like you, they can shut you down. If you use an open model like R1, you always have the option to self host even if it can be costly, and not be at the mercy of a third party being able to just kill your business by shutting down your access to their service.
If that's an issue, there's a workaround using structure generation to force it to output a </thiking> token after some threshold and force it to write the final answer.
> And where does the nuclear fuel come from? Russia.
Not true at all. Russia is producing 5% of the world Uranium, and they probably use quite a lot of that domestically given they produce 8% of all nuclear power in the world with their own plant.
Kazakhstan + Uzbekistan is 50% of the word production. Canada is second and will be happy to start selling to the EU. Namibia and Australia both produce twice as much as Russia.
Not to say that supply of natural Uranium is not a concern because you do depends of a small list of countries but we don't need to buy any from Russia.
The problem with the original Kinect (v1) is that good tracking software for it was never really written. Most application that support it, just use the original Microsoft SDK to do the motion tracking, and it's just not very good, the main issue is that it always assume that the tracked person is directly facing the camera, and is very bad at dealing with occlusion. The good thing about it, is that it ran in real-time on a potato.
In order to get a good result, someone would need to train a good model for HPE that could use the cloud point data directly, but it seems nobody cares about depth sensor anymore, most efforts are going to HPE from regular 2d video (like media pipe holistic model). And given the result you can get with media pipe, openpose and the likes, it's understandable nobody is bothering working with low resolution cloud point anymore for 3D HPE.
The only use-case I can think of for a Kinect v1 in 2025, would be robotics if you want a low latency low resolution cloud point for your robot control, but even there I think we are moving to big vision model capable of making sense of regular video feeds.
Saying this is similar to Moshi is like saying GPT2 is similar to GPT4. You can't have any sort of conversation longer than 30s with moshi before it goes banana. You can talk to this model for an hour and it remains completely coherent.
At first, write clean code with functions and don’t obsess over call overhead. Once it works, profile, then optimize where it actually matters.
Premature optimization, etc.