> If it was entirely up to fable max or sol max the result would have been pretty bad.
How can you know it will have failed?
I don't think it's that hard, if you clearly define the goal well, and have a bit more compute available, and do some intermediary bookkeeping.
Been running near identical tests for years now. Latest models are the only ones I don't throw away the results/code. Which is impressive, salvagable/usable is a giant step up.
I'm thinking of the Ottomans capturing Constantinople. Very much paraphrasing: Since Ottomans had a massive cannon made with new technology. Since Constantinople hadn't hired the cannon inventor. Since Constantinople was poor. Since it was sacked by other christians in the crusades previously. So it was self afflicted by christians.
One can see the climate crisis like that. Only global accords work, otherwise self-limiting CO2 is mostly sacrificing one's economy while others do better. USA will block all international climate accords. Because the president and half of politicians acts like climate change is not real. Because a whole party thinks that. Because voters think that. Because the media tells them that. Because the media treats fossil company propaganda like science. Because it's so cheap for a handful of companies to corrupt media or politics.
What is the epistemology of the majority of voters... scandals, conspiracy theories, UFO:s. What shows on TV or on your feeds. I assume it's child's play for some non-western actor to cause constant infighting in the west, causing billions in economic losses or destruction of alliances. For whole planetary level issues, who's the opponent, working in the shadows? Maybe, after all, the aliens are already here...
Sometimes I feel the same, and then sometimes I wonder if that "moving away feeling" is the last of the old breed who are hanging onto power, maybe the next step after this massive wave of ultra right wing stupidity is something much brighter ?
The western world has more or less come together to reduce emissions. If you look at US for instance emissions are down in abs terms. Not relative to gdp or per capital but total emissions. This holds even if you adjust for trade.
The “problem” is there will always be independent upcoming nations like china and India that don’t really care to make sacrifices for collective goals. Their people are finally experiencing materially better lives and that comes w energy consumption.
This will ultimately be solved by markets and technology. People 100y ago were concerned about horse crap flooding the streets and what solved that wasn’t some agreement to limit horses but the invention of the automobile. Same will happen here and it’s already beginning to happen
> The western world has more or less come together to reduce emissions. If you look at US for instance emissions are down in abs terms. Not relative to gdp or per capital but total emissions. This holds even if you adjust for trade.
It certainly is reduced per-production. China has double the carbon emission intensity of production that the USA has for example.
I'm not sure China's emissions can be considered a separate thing when the rest of the world relies on them for plenty of manufacturing. The West is still a massive emitter per capita.
Terrence Tao is an exception in many regards. I’m less concerned if it helps him, and more interested in how it affects the average person, since the vast majority of users will be average.
Tao can use LLMs on his own terms, on his conditions. Under those circumstances, it's an absolutely fantastic tool.
In an enterprise, everyone gets to use LLMs. Everyone can create slop. In an enterprise environment, this can lead to weird situations where the senior has to put in extra efforts to contain the slop of others... so I can kind of see where the "hinders seniors" is coming from.
I don’t think it’s even “seniors” containing slop of others.
I think it’s those who still care about quality (which seems to be decreasing in the industry overall) trying to keep the whole thing running, in a sea of people who think going really really fast and not understanding their code is fine because LLMs can make sense of it and clean it up later.
Just look at the condition of open source contributions recently
Well, they might have gotten a little wary from previous boom and bust cycles.
Perhaps they are a bit wary about the economic sustainability of the whole AI thing.
However, perhaps they also might be driven by greed at this point. Why not just constrain supply and increase margins whilst they are no real competitor?
> TSMC’s leadership dismissed Altman as a “podcasting bro” and scoffed at his proposed $7 trillion plan to build 36 new chip manufacturing plants and AI data centers.
I thought it was ridiculous when I read it. I'm glad the fabs think he's crazy too. If he wants this then he can give them the money up front. But of course he doesn't have it.
After the dot com collapse my company's fabs were running at 50% capacity for a few years and losing money. In 2014 IBM paid Global Foundries $1.5 billion to take the fabs away. They didn't sell the fabs, they paid someone to take them away. The people who run TSMC are smart and don't want to invest $20-100 billion in new fabs that come online in 3-5 years just as the AI bubble bursts and demand collapses.
I started working during the dot com boom. I was getting 3 phone calls a week from recruiters on my work telephone number. Then I saw the bubble burst in from mid-2000. In 2001 zero recruiters called me. I hated my job after the reorg and it took me 10 months to find a new one.
I know a lot of people in the 45+ age range including many working on AI accelerators. We all think this is a bubble. The AI companies are not profitable right now for the prices they charge. There are a bunch of articles on this. If they raise prices too quickly to become profitable then demand will collapse. Eventually investors will want a return on their investment. I made a joke that we haven't reached the Pets.com phase of the bubble yet.
I've thought about this too.
I do agree that open source models look good and enticing, especially from a privacy standpoint.
But these solutions are always going to remain niche solutions for power users.
I'm not one of them.
I can't be hassled/bothered to setup that whole thing (local or cloud) to gain some privacy and end up with an inferior model and tool. Let's not forget about the cost as well!
Right now I'm paying for Claude and Gemini.
I run out of Claude tokens real fast, but I can just keep on going using Gemini/GeminiCLI for absolutely no cost it seems like.
The closed LLMs with the biggest amount of users will eventually outperform the open ones too, I believe.
They have a lot of closed data that they can train their next generation on.
Especially the LLMs that the scientific community uses will be a lot more valuable (for everyone).
So in terms of quality, the closed LLMs should eventually outperform the open ones, I believe, which is indeed worrisome.
I also felt anxious early december about the valuations, but, one thing remains certain.
Compute is in heavy demand, regardless of which LLM people use.
I can't go back to pre-AI. I want more and more and faster and faster AI.
The whole world is moving that way it seems like.
I'm invested into phsyical AI atm (chips, ram, ...) whose evaluations look decently cheap.
I think you should reconsider the idea that frontier models will be superior, for a couple reasons:
- LLMs have fixed limitations. The first one is training, the dataset you use. There's only so much information in the world and we've largely downloaded it all, so it can't get better there. Next you can do training on specific things to make it better at specific things, but that is by definition niche; and you can actually do that for free today with Google's Tensors in free Cloud products. Later people will pay for this, but the point is, it's ridiculously easy for anyone to fine-tune training, we don't need frontier companies for that. And finally, LLM improvements come by small tweaks to models that already come to open weights within a matter of months, often surpassing the frontier! All you have to do is sit on your ass for a couple months and you have a better open model. Why would anyone do this? Because once all models are extremely good (about 1 year from now) you won't need them to be better, they'll already do everything you need in 1-shot, so you can afford to sit and wait for open models. Then the only reason left to use frontier cloud is that they host a model; but other people do cloud-hosted models! Because it's a commodity! (And by the way, people like me are already pissed off at Anthropic because we're not allowed to use OAuth with 3rd party tools, which is complete bullshit. I won't use them on general principle now, they're a lock-in moat, and I don't need them) There will also be better, faster, more optimized open models, which everyone is going to use. For doing math you'll use one model, for intelligence you'll use a different model, for coding a different model, for health a different model, etc, and the reason is simple: it's faster, lower memory, and more accurate. Why do things 2x slower if you don't have to? Frontier model providers just don't provide this kind of flexibility, but the community does. Smart users will do more with less, and that means open.
On the hardware:
- Def it will continue to be investment-worthy, but be cautious. The growth simply isn't going to continue at pace, and the simple reason is we've already got enough hardware. They want more hardware so they can continue trying to "scale LLMs" the way they have with brute force. But soon the LLMs will plateau and the brute force method isn't going to net the kind of improvements that justify the cost. Demand for hardware is going to drop like a stone in 1-2 years; if they don't cease building/buying then, they risk devaluing it (supply/demand), but either way Nvidia won't be selling as much product so there goes their valuation. And RAM is eventually going to get cheaper, so even if demand goes up, spending is less. The other reason demand won't continue at pace is investors are already scared, so the taps are being tightened (I'm sure the "Megadeal" being put on-hold is the secret investment groups tightening their belts or trying to secure more favorable terms). I honestly can't say what the economic picture is going to look like, but I guarantee you Nvidia will fall from its storied heights back to normal earth, and other providers will fill the gap. I don't know who for certain, but AMD just makes sense, because they're already supported by most AI software the way Nvidia is (try to run open-source inference today, it's one of those two). Frontier and cloud providers have Tensors and other exotic hardware, which is great for them, but everyone else is gonna buy commodity chips. Watch for architectures with lower price and higher parts availability.
> There's only so much information in the world and we've largely downloaded it all, so it can't get better there.
What about all the input data into LLMs and the conversations we're having?
That must be able to produce a better next gen model, no?
> it's ridiculously easy for anyone to fine-tune training, we don't need frontier companies for that.
Not for me. It'll take me days, and then I'm pretty sure it won't be better than Gemini 3 pro for my coding needs, especially in reasoning.
> For doing math you'll use one model, for intelligence you'll use a different model, for coding a different model, for health a different model, etc, and the reason is simple: it's faster, lower memory, and more accurate.
Why wouldn't e.g. Gemini just add a triage step?
And are you sure it's that much easier to get a better model for math than the big ones?
I think you underestimate the friction this causes regular users by handpicking and/or training specific models, whilst the big vendors are good enough for their needs.
> What about all the input data into LLMs and the conversations we're having? That must be able to produce a better next gen model, no?
Better models are largely coming from training, tuning, and specific "techniques" discovered to do things like eliminate loops and hallucinations. Human inputs are a small portion of that; you'll notice that all models are getting better despite the fact that all these companies have different human inputs! A decent amount of the models' abilities come from properties like temperature/p-settings, which is basically introducing variable randomness. (these are now called "low" and "high" in frontier models) This can cause problems, but also increased capability, so the challenge isn't getting better input, it's better controlling randomness (sort of). Even coding models benefit from a small amount of this. But there is a lot more, so overall model improvements are not one thing, they are many things that are not novel. In fact, open models get novel techniques before the frontier does, it's been like that for a while.
> Not for me. It'll take me days, and then I'm pretty sure it won't be better than Gemini 3 pro for my coding needs, especially in reasoning.
If you don't want the improvements, that's up to you; I'm just saying the frontier has no advantage here, and if people want better than frontier, it's there for free.
> Why wouldn't e.g. Gemini just add a triage step? And are you sure it's that much easier to get a better model for math than the big ones?
They already do have triage steps, but despite that, they still create specific models for specific use-cases. Most people already choose Thinking by default for general queries, and coding models for coding. That will continue, but there will be more providers of more specific models that will outperform frontier models, for the simple fact that there's a million use-cases out there and lots of opportunity for startups/community to create a better tailored model for cheaper. And soon all our computers will be decent at doing AI locally, so why pay for frontier anyway? I can already AI-code locally on a 4 year old machine. Two years from now, there likley won't be a need for you to use a cloud service at all, because your local machine and a local model will be equivalent, private, and free.
I agree.
Especially the whole Johny Ive and Altman's hype video in that coffee shop was absolutely disgusting. Oh how far their egos have been inflated, which leads to very bad decision making. Not to be trusted.
Google can spy on everything: via its OS, its browser, its Youtube, its search engine, its ad network, its blog network, its maps app, its translation service, its fonts service, its 8.8.8.8, its Office suite, its captcha, its analytics service, and on and on and on...
My brother convinced me to try a 1Password family account, since it would be cheaper. Ever since, the Chrome plugin takes forever to login. Sometimes up to 5-6 seconds. And it really annoys me that they have so many resources and money, and it's still this expensive for a very very basic application, and slow to boot.
I tried out passwords, and combined with Safari, it's an absolute godsend compared to 1Password.
That does mean that I switched from Brave to Safari, and thus have YouTube ads, and so I'm now paying for YouTube haha
> Ever since, the Chrome plugin takes forever to login.
This isn't my experience since the recent update that shows up a mini-login panel when trying to sign in. The old experience that opened the desktop app first was fairly slow.
Just to rub it in your face :) (teasingly and with respect) I got Android/LastPass/Firefox and only pay for the LastPass annually (I got it on all my devices), so there you have it ;)
Does it blank all the fake videos in your YouTube home page? There used to be ads separate. Then they started putting one in the upper left corner that pretended to be a real video, with some clickbait title. Now (today?) they have them sprinkled all over, like maybe 15% of all the thumbnails are now ads.
I'm leaving that platform. They've taken shittification to new heights.
At the root of medicine is biochemistry, and at the root of chemistry is quantum physics: the formation and breaking of chemical bonds is at core random events shaped by probabilities. We can only say how likely events are, not which ones will happen when.
How can you know it will have failed? I don't think it's that hard, if you clearly define the goal well, and have a bit more compute available, and do some intermediary bookkeeping.
reply