> But the code quality is speed. And reach. You can not advance, unless you can read the code, you can understand the model, you can not scale beyond a certain point
Other people can do the important work of investing time to understand the model and simplify the code architecture, as proven many times over by actively maintained projects pioneered by Fabrice.
To kickstart a project, you have to show people that something they assumed impossible or hard to achieve is actually possible by dropping it in front of them.
> Its a great interview topic to filter this kind of candidate out of companies.
Fabrice Bellard ships. It makes sense to filter him out if you're a bank or an org with well-established products that prefers stability over velocity. If you're a start-up or have lots of greenfield projects requiring fast experimentation loops: you need folk who can ship quickly. Most organizations have a mix of projects and need a healthy mix of engineers, or ones who can flip modes relevant to the project.
Indeed, the American proletariat need not seize the means of production through local AI. Instead, full control of AI ought to be left in the benevolent hands of the billionaire capitalist class, and the benefits will trickle down, perhaps some time after after the 100th sponsored study into UBI (which very different from socialism)
> Sure. Legally, makes sense. Practically, if you want to do all those things...
If you had just owned up to how you were mistaken about EU legislative limits - confidently stated - I probably would have taken everything else in your initial comment at face value.
Your doubling down into unfalsifiable territory has me thinking your arguments are feelings-based with post-facto justifications.
> how you were mistaken about EU legislative limits
I’m not making any legal arguments. The fact that the EU can’t legislate on those issues doesn’t change that its AI Act has those loopholes.
> unfalsifiable territory
No, I’m not. If the AI Act constrained any actual risks, that would falsify my assertion. I’m saying it in practice doesn’t. Those capabilities are still being built, just not in Europe. And they’ll still be sold to Europe, just to its governments to use however they want, not to its people.
The EU doesn’t have the power to write AI legislation for human rights purposes. It does have the power to throw gum into its AI industry’s works. It did what it could. Which is very little of the former (by constraining B2C and B2B, sort of). It did a lot of the latter.
Congress can’t do a lot of things. Passing something stupid and then complaining that the reason it isn’t competently written is because of Constitutional limits doesn’t absolve the stupid bill.
I’m not an expert on EU law or AI. But I do make capital-allocation decisions around this stuff, and I know enough to know that as currently configured the only main AI business to do in the EU is in selling it things that kill or surveil.
I read their post in the way they intended. Regardless of whether they can, the fact that they fail to cover all the bases makes the legislation almost useless.
> In this case though, I'll admit that it would be a negative signal if they never tried it even once and refuse to do so
I was arguing that this is a bad question elsewhere but you provided another reason. If a candidate tells you they haven't tried using AI (without saying why), that offers no signal at how well they'll do if you hire them, and you construe it as a negative signal. If you want to know if they would be willing to use AI as part of the job ask that question instead!
> You can't make a solid opinion on things you never try after all.
In this case, I wouldn't mind if they used AI or not. Honestly it'd be quite interesting to find someone who doesn't use it to hear about why. At the end of the day, I'm more concerned about how effective they'd be in their role. And if I have 2 candidate, one with better domain knowledge but no AI, and another with AI but no domain knowledge, I'd most likely pick the one with domain knowledge.
And if they can make an argument about why they're more effective without AI, more power to them. So if I were to ask the question directly as you suggest, it would be: Are you more effective than the other person I'm interviewing? - which i guess is the whole point of the interview.
> Heroin
Fair point, but I also don't think that trying AI is anywhere close to the level of trying Heroin
> The most impressive think that stuck with me is that humans are incredibly efficient, from an energy perspective, in anything we do, compared to machines.
Humans are efficient, but not across the board. Trivial counterexample: walking is incredibly energy inefficient vs a bicycle or other wheeled conveyances whose primary dissipater is rolling resistance.
We're still pretty efficient while not having wheel shaped limbs. Running like humans works pretty well. So well even that we can chase a lot of animals longer than they can outrun us.
There might be more efficient ways to move but we are pretty well equipped by evolution.
It's not strange at all, I was responding to a specific, incorrect claim. I even quoted the wrong claim in my earlier comment , and I'll repeat it again, with added emphasis
>>> humans are incredibly efficient, from an energy perspective, in anything we do, compared to machines
I simply provided contrary evidence to a well-defined, falsifiable claim. How is that strange?
Yes, but walking and moving on wheels is oranges and apples. It would be a relevant comparison if a robot with a movement mechanism based on two feet was more efficient than a human.
> in one assignment I remember comparing the energy outputs between the human and robot equivalents of different tasks, whether or not the robot was humanoid in how it was designed
So I think the point in this context is relevant, even if it's apples to oranges.
The point isn't that a humanoid robot walking is less efficient than a human walking, is that moving on a wheel is not the same thing as walking. For example, using wheels is not only less efficient it is barely usable for climbing rocks, going up the stairs and many other surfaces that makes the comparison irrelevant.
You could say that a robotic gun is much more efficient than a human in killing, that's another easy easy comparison of different tasks where robots win, but it totally miss the point.
I’ll admit, at first, I thought the human vs machine comparison was about humanoid machines. But that’s too narrowly defined to be a useful comparison. Most machines in use today are not humanoid.
Then to boldly claim that humans are more efficient at anything compared to a machine, just does not follow.
> On the other hand, even if Apple's AI were 6 - 9 months or a generation behind,
Do you mean Google's AI with Apple wrappers? Apple's in-house AI is further behind Google, amd very far from the frontier according to your ranking. IMO, Google is on the frontier - I recall Altman calling for an OpenAI all-hands-on deck when Gemini was released because of how good it was compared to ChatGPT. I also suspect Google has the lowest operating expenses due to scale, experience and luck/planning (TPUs), there will come a time when AI investments will slow down, and the cost of revenue will become more important.
> ...the research progresses before the inevitable nationalization of the frontier.
Hacker News has been telling me America beats China at "innovation" because of the "freedoms" - especially frew enterprise. I wonder how a nationalized frontier lab would perform.... Andhow the non-citizen researchers would feel about working for the US government that doesn't trust them to use frontier models.
Model effectiveness has improved across model sizes. You really should try the latest flash variants more. They have become my default for most tasks except for gnarly high-level planning.
"Capability per parameter" is rising, but parameter count remains an advantage. And small models remain bad, because "good" is a rapidly moving target.
A 2026 4B beats 2024 4B, but both are far behind the contemporary frontier. Which makes them bad. There is no such thing as "too much capability" - a "good" model is whatever the current frontier is.
In 2024, a "good" model is one that can be trusted to write a 800 line script. In 2026, it's a model that can be trusted to do gnarly high-level planning and execution both. In 2028, it's going to be something like a model you can point at an extremely involved task, abandon, and have it report back with a "done" in 3 weeks.
> A 2026 4B beats 2024 4B, but both are far behind the contemporary frontier.
The thing about engineering is you don't just use the biggest bolt on the market on every bridge.
> In 2024, a "good" model is one that can be trusted to write a 800 line script. In 2026, it's a model that can be trusted to do gnarly high-level planning and execution both
This sounds a lot like having a single diamond-head hammer as the only tool in your toolbox. As suggested by the name, flash models are fast - sometimes I want to write the equivalent of fifty 800-line scripts. There is such a thing as good enough.
Good enough? That's a lie people tell each other because they lack imagination.
"It's good enough" was said about GPT-4, o1, o3, Opus 4 and more. Guess what happened? Newer models released, people updated their expectations of what LLMs can do, usage got more aggressive, and somehow, GPT-4 went from "good enough" to "obsolete trash".
If you have no imagination, then at least substitute your pattern recognition for it.
The world is hungry for capabilities. There are piles upon piles of tasks that aren't done by LLMs simply because LLMs aren't good enough to do them.
The thing a frontier model gives you is "you don't have to babysit a model to get it to do X", and that X gets more and more impressive release to release.
I wish you had addressed at least one of arguments in good faith before jumping to insults and countering a strawman argument I didn't make - I never claimed their will be no use for more capable models.
You do your AI-maximalism, and I'll stick to making trade-offs based on the needs of each piece of work.
Right - the idea that "bigger model = better" might have been true a year ago, but the flash models are extremely effective right now. You simply use them for the tasks they are ideally suited for.
You call the response "cagey and evasive", but that is for an objectively a bad interview question, one wrung below "How many years experience do you have prompting Anthropic Opus? We are an Opus shop." People are not locked into their current way of using AI and it is trivial to match how one works with AI to match employers requirements. It's a question that deserves an idealized non-answer
Remember the context - this is while solving a whiteboard problem. Its bad in the same way asking candidate what their birthstone is - because any answer offer little or no signal about the odds of a candidates success at the company.
I'm curious to know why you think asking about AI usage is a good interview question.
If you're going to have people use AI regularly, it's worth asking so that you can get a sense of their interest/willingness, experience level, and training needs. That said, more specific questions are typically more revealing. Personally I'm fond of "If you could give Claude only 1 instruction, what would it be?"
Other people can do the important work of investing time to understand the model and simplify the code architecture, as proven many times over by actively maintained projects pioneered by Fabrice.
To kickstart a project, you have to show people that something they assumed impossible or hard to achieve is actually possible by dropping it in front of them.
> Its a great interview topic to filter this kind of candidate out of companies.
Fabrice Bellard ships. It makes sense to filter him out if you're a bank or an org with well-established products that prefers stability over velocity. If you're a start-up or have lots of greenfield projects requiring fast experimentation loops: you need folk who can ship quickly. Most organizations have a mix of projects and need a healthy mix of engineers, or ones who can flip modes relevant to the project.
reply