The thing is that this thing is constantly compacting.... I get 1M context with Claude and ~256k with Astra. Even if the compaction loses much less information on OAI's side, it takes so long it's barely any use for me...
I've tried High and Max. They have produced decent results, but they're so slow.... I will try to lower it a bit and see the difference, but it's a delicate balance: I don't want to waste literal hours on the incorrect reasoning level to only then have to spend those hours and tokens to do it right.
At this very moment, Astra has been working for 1h15m on a task. At this rate I genuinely expect it to take about 10 hours. I feel like claude would do it in at least a third of that. Let's see if the quality justifies the slowness (it better)
It should burn N + len(answer), because you have to re-cache the whole answer without the prompt stack.
Perhaps more persnickety, it pushes the LLM out of distribution - if it’s unnatural for it to write in plain language without the prompt stack, your prefix will be an unnatural conversation which can reduce intelligence in hard to measure ways, especially over long conversations.
Not saying don’t do it, clarity is perhaps worth the intelligence hit, but it’s not going to be a free lunch.
Yes, maybe. Evidence around caveman shows this isn’t a big deal for token consumption (forcing it to respond in a way it was not tuned) and I don’t see a big difference either way. And maybe it increases intelligence in hard to measure ways. Lots of parameters in these models.
I feel I fight them less with this setup. They get so lost in their own invented bullshit they stop being useful pretty often without it. So I would take bets on that :)
“Users are the product” is a phrase used when the users aren’t the ones paying for a free service. For a paid API the users absolutely are the customers.
FOSS is the wrong analogy. Building frontier LLMs isn’t primarily an engineering discipline, it’s a scientific research program.
Of course we do have basically open source research programs, including most universities and big projects like CERN. But AI grew up in universities until it transpired that sufficient capital could only be found in the private sector.
It would be possible to make a decent publicly funded AI research program. But it would look more like the Manhattan or Apollo projects (which frontier labs already model themselves after) than some extra research grants for universities.
The Manhatten project cost about $40 billion in total adjusted for inflation. Anthropic's latest funding round alone raised $65 billion.
The entire Apollo project at the peak of the cold war cost about $300 billion in today's dollars. That's approximately what OpenAI and Anthropic have raised together in total until now.
I don't think governments can supply this amount of money for AI in the current political and economic climate. The LHC cost less than $10 billion by comparison and it was spread out over a much longer timeframe.
> I don't think governments can supply this amount of money for AI in the current political and economic climate.
I'm a believer in Keynes' "anything we can do we can afford". It could be afforded .. if there was a sufficiently good reason. And there isn't. This is way behind "governments, especially the EU, should have a sovereign cloud". It is also way behind "governments need to keep global warming below 2C by the end of the century" and "governments need to ensure affordable energy", objectives which the current AI buildout is in direct conflict with.
This is before we get into the question of whether AI has net positive social value in non-software use cases. Even in software the case for AI is explicitly job-destroying and raising electricity prices for everyone else.
The only reason AI research is in conflict with limiting global warming is the unholy alliance of NIMBYs and Sierra Club environmentalists preventing the building of anything, anywhere. Clean energy has beaten out fossil energy on cost for years now, and only regulatory burdens force datacentres to use portable gas turbines instead of standing up transmission lines to the closest solar or wind farm.
There are things you can't do - even as a society - when the costs start ballooning beyond the GDP of smaller European nations. I mean, yes, in an emergency you could cancel all social security and public infrastructure spending and instead dump it into AI. But even then it will take several big nations to foot the bill if you want to still have a country left after you're done. With the level of political heckle everywhere, this is just not going to happen, because it would need unanimous support from all sides.
What you are potentially missing is the public sector wouldn't necessarily be in the same race as the commercial companies.
It doesn't matter if a public effort is a year or two behind the curve ( and hence has dramatically lower costs due to Moores law and the ability to piggy back on research ) - especially as we approach the asymptotic phase of development.
> I don't think governments can supply this amount of money for AI in the current political and economic climate.
You understand how the system works if you’re thinking in terms of government/non-government. The current political and economic state is not a bug, it’s by design which serves a purpose.
Yes, if the US was modeled like the USSR they could extract 100% of everyone’s private assets and the politburo could spend them on whatever moonshot projects they wanted to, unchecked by silly democratic votes.
But…the USSR and that entire model failed spectacularly? So not sure what you’re getting at here. Is there some fantasy economic model you believe you’ve innovated that will lead to utopia and the end of resource scarcity?
Govts can take it over. Corporations dont maintain standing armies. So there is a pecking order that corps have never been able to invert. They rely on Govt for their own security.
History is full of these take overs if there is risk(usually happens after some catastrophe). See the finance sector(once upon a time private banks invented and printed money), nuclear industry, febrtilizer industry, crypto, a whole bunch of processes in biotech/synthbio. Classic textbook example is the East India Company. It was much richer that the British Govt or the King.
> But AI grew up in universities until it transpired that sufficient capital could only be found in the private sector.
One way of looking at it. Another is that AI research progressed within universities, but it was only until recently that the private sector saw the profit potential when combined with modern CPU/GPU technology.
You could argue we're saying the same thing, but I think these angles are different. An academic research programme would not have spent billions on a datacentre to provide AI for free to the general public, for example.
I agree an academic research program wouldn’t spend billions on a datacenter, but that’s just a reason why foss AI as described in this article won’t work - because big compute is required to do anything of real relevance.
Very interesting! I honestly would have expected the opposite: I was optimistic about strongly functional languages in the age of AI. The more modular and side-effect-free your code is, the easier it should be to constrain changes, catch slop, and reduce LLMisms like spooky hacks-at-a-distance.
It looks like the issues are in the compiler and documentation so hopefully it’s fixable… I write in Python every day but I do miss smarter languages and I hope AI doesn’t fully obliterate them.
The next Jonathan Blow is going to be massively empowered by their tools and make something wonderful. Having fewer people involved can lead to a more focused execution of their vision - most amazing indie games are like this. But yes your average game isn’t bad because it’s hard to write C#, it’s bad because it’s hard to design great unique mechanics and levels, and it’s hard to see AI helping (indeed not harming) that.
Let LLMs help you with coding. Design the game and the mechanics yourself. I can see this being an incredibly empowering tool in the right game developer's hands; but if you come into it with a token-maxxing / AI-maxxing mentality, I doubt you'll make a fun game to play.
I love the concept - blocking apps are often too restrictive which makes me disable them. Slowing could be a nice alternative.
This probably uses a vpn? It’s important to think about how to stop me disabling it casually. I use Opal which blocks my settings page too. Which works great but frustratingly it blocks my legitimate needs very often too!
reply