> I get that they rely on people not using all of their quota
They have no problem with users using their quota on their own software. Because they get the signals. They do have a problem with users using the API in 3rd party software, because they don't get the signals.
That's very different from what I'm seeing around me, but yes, I suppose that happens to. And I guess Google wouldn't have as much of an issue with that, right?
Ah, in my spaces (Involved in the proxy dev), most people have been using it for Opus. I suspect they may even have more of an issue with it, as they don't get the cost advantage of serving an in-house model
I don't really understand this reasoning actually:
if OpenClaw usage go up, and a service (OpenAI it looks like) gets lots of usage data for personal assistent usage, they can optimize to make it better for people who get a $200 subscription just because of that use case.
I think it's one peg below intel agencies. It's the local gov agencies that want that power. The 3 letter peeps can already tell who writes what, both at scale and targeted.
> and we'll be able to see if all the hype was warranted.
Umm, what? For the past 3 years, every year I've said something along the lines of "even if models stop improving now, we'll be working on this for years, finding new ways to use it and make cool stuff happen". The hype is already warranted. To have used these tools and not be hyped is simply denial at this point.
Maybe AI is useful to you, but the US economy is currently buoyed by promises of AI replacing the workforce across the board.
Most of Mag-7 are planning to spend over 500B on capex this year alone on building out datacenters for AI pipelines that have yet to prove that it can generate a sustainable profit. Yes, AI is useful in some environments, but the current pricing is heavily subsidized. So my point stand, the hype is not warranted.
> but the US economy is currently buoyed by promises of AI replacing the workforce across the board.
Still don't understand what's the end goal here. Assuming they don't deliver, then there are billions of investments that will go bust. Assuming they deliver, millions lose their jobs and there's going to be a bloodbath on the streets.
the end goal is productivity growth, aka the point of nearly every technology ever invented. The human story is about how we learn to do more with less.
> Assuming they don't deliver, then there are billions of investments that will go bust. Assuming they deliver, millions lose their jobs and there's going to be a bloodbath on the streets.
There is a third outcome that combines both of these.
LLMs can massively displace the workforce (and cause widespread social instability) AND the companies pouring hundreds of billions into them right now could, at the same time, fail to capture significant amounts of the labor savings value as late-mover alternatives run the race drafting their progress without the massive spend.
I'd honestly be surprised if this double-whammy isn't the outcome at this point. AI is going to have a massive impact on everything, but there is still no moat in sight.
Leaving aside the economic shitshow and other things.
I think you're right but for the wrong reasons wrt sustainable profit.
Specifically, overcounting how much it will cost in 5 years to run AI because you're extrapolating current high prices, and at the same time undercounting how the demand will drive efficiency gains.
I think our little corner of the world has a distorted view of AI in that it is actually proving useful for us. Once they passed a certain level of usefulness... I remember when they were still struggling just to output syntactically correct code, you know, like, 18 months ago or so... they became a useful tool that we can incorporate.
But there's a lot of things playing out to our advantage. Vast swathes of useful and publicly available training data. The rigorous precision of said data. Vast swathes of data we can feed it as input to our queries from our own codebases. While we never attained the perfect ideal we dreamed of, we have vast quantities of documentation at differing levels of abstraction that the training can compare to the code bases. We've already been arguing in our community about how design patterns were just level of abstraction our coding couldn't capture and AI has access now to all sorts of design patterns we wouldn't have even called design patterns because they still take lots of code to produce, but now for example, if I have a process that I need to parallelize it can pretty much just do it in any of several ways depending on what I need at that point.
It is easy to get too overexcited about what it can do and I suspect we're going to see an absolute flood of "We let AI into our code base and it has absolutely shredded it and now even the most expensive AI can't do anything with it anymore" in, oh, 3 to 6 months. Not that everyone is going to have that experience, but I think we're going to see it. Right now we're still at the phase where people call you crazy for that and insist it must have been you using the tool wrong. But it is clearly an amazing tool for all sorts of uses.
Nevertheless, despite my own experiences, I persist in believing there is an AI bubble, because while AI may replace vast swathes of the work force in 5-20 years, for quite a lot of the workforce, it is not ready to do it right this very instant like the pricing on Wall Street is assuming. They don't have gigabytes of high-quality training data to pour in to their system. They don't have rigorous syntax rules to incorporate into the training data. They don't have any equivalent of being guided by tests to keep things on the rails. They don't have large piles of professionally developed documentation that can be cross-checked directly against the implementation. It's going to be a slower, longer process. As with the dot-com bubble, it isn't that it isn't going to change the world, it is simply that it isn't going to change the world quite that fast.
i think the point is AI has to go much further and faster than it has in the past 3 years to justify the investments being made from the hype. The hype did its job now the AI industry has to execute and create the returns they promised. That is still very much up in the air, if they can't then the tech was over hyped.
It's high time to stop accumulating debt while providing free picture of pelicycles, just charge the full cost for them - enough to generate profits and pay back debt.
What we see now is literally burning money and energy to generate hype. The only true measures of success are financial and macroeconomic. If the hype is real, there should be no problem for the mighty AI to generate debt-free profits for its providers while the overall price level in the US goes down.
We observe the exact opposite which makes the AI hype act only as market manipulation for capital misallocation.
unlike the old hpc, where we only burned hundreds of millions for machines that were 80% efficient to get a 5 year lead, we are burning hundreds of billions on machines that are 30% efficient to get a 1 year lead.
I don't have inside info, but everything we've seen about gemini3.0 makes me think they aren't doing distillation for their models. They are likely training different arch/sizes in parallel. Gemini 3.0-flash was better than 3.0-pro on a bunch of tasks. That shouldn't happen with distillation. So my guess is that they are working in parallel, on different arches, and try out stuff on -flash first (since they're smaller and faster to train) and then apply the learnings to -pro training runs. (same thing kinda happened with 2.5-flash that got better upgrades than 2.5-pro at various points last year). Ofc I might be wrong, but that's my guess right now.
Interesting. Whatever they are doing it's a bit different than Anthropic and oAI, which is good for the consumer. I'm curious about their ML Ops internally; would be fascinating to learn more.
> I think the premise of this tool is flawed. Bad faith actors are not people who...
I think there's something here. The tool is not intended to stop bad faith actors. You can't stop those. But you can nudge people into "being better" with a simple prompt. I can't recall the exact blog/paper now, but I remember reading that someone did this test (google perhaps?) and saw that with a simple prompt "hey this message is high on anger, did you mean to write it like this?" before submitting lead to ~30-50% to change their message and tone it down. It might help in that regard.
Case in point IMO, this very comment embodies the common trap that even "smart" people fall into... black-and-white thinking that is often wrong and just makes you look less intelligent.
"As a rule, strong opinions about issues do not emerge from deep understanding."
That sub used to be the absolute best place to get the latest in LLM developments. The worst thing that happened to the sub was karpathy making it popular with a tweet. Since then it's been overrun by a whole bunch of drama, toxic behaviour and useless bots, and the quality content has cratered.
There was a mod crisis and new mods came in, with really weird stuff (integrations with discord and such), lots of bots became active with useless posts and "engagement" bait, the chinese labs are all fighting eachother on who's better every time there's a release, claude-induced-manias on "papers" this and "zenodo" that (everyone is a researcher now, everyone is inventing a subquadratic attention, led by claude hallucinated stuff), they have an obsession with "local only", leading to removing any discussion about SotA (which is entirely counter productive) and so on.
Isn't that a bit too certain for something that's not settled at all? How else would you explain the Polgar sisters? I'm sure there are other examples, but this is the most famous one.
Few claims in the social sciences are more fully settled. I don't think you could find a researcher in the world at a major university making that claim that randomly selected children could be reliably turned into world-class mathematicians with enough training.
> How else would you explain the Polgar sisters?
Genetics is the obvious explanation. The father was clearly very intelligent.
Also to clarify: I agree that training and effort can have large effects, and that focusing on them is a good strategy. Over-believing in them is probably a good bias, even. But the idea that everyone is more or less the same except for effort is ridiculous.
Yep - requires the client to trust the SSL cert of the proxy. Cooperative clients that support eg HTTP_PROXY may be easier to support, but for Airut I went for full transparent mitmproxy. All DNS A requests resolve to the proxy IP and proxy cert is injected to the container where Claude Code runs as trusted CA. As a bonus this closes DNS as potential exfiltration channel.
Anthropic have been the loudest in pushing for regulatory capture, often citing "muh security" as FUD. People should care what they write on this topic, because they're not writing for us, they're writing for "the regulators". Member when the usgov placed a dude in solitary confinement because they thought he could launch nukes with a whistle? Yeah... Let's hope they don't do some cray cray stuff with open LLMs.
Anthropic make amazing coding models, kudos for that. But they should be mocked for any communication like the one linked. Boo-hoo. Deal with it, or don't, I don't care. No one will feel for you. What goes around, comes around. Etc.
Administratively, Anthropic seems to misunderstand politics. You don't get to wear the "people's champion" and "government sweetheart" hats at the same time, when push comes to shove you'll be forced to pick a lane. We saw it with Microsoft, we saw it with Apple and Google, and now we're seeing it with OpenAI too. You can't drive down both paths at the same time.
As a member of the target audience for Claude, their messaging just leaves me confused. Are you a renegade success, or do you need the government's help? Are you a populist juggernaut, or do you hide from competition? OpenAI, for all their myriad issues, understood this from the start and stuck to the blithely profitable federal ass-kisser route.
They have no problem with users using their quota on their own software. Because they get the signals. They do have a problem with users using the API in 3rd party software, because they don't get the signals.
reply