No, but I wouldn't be able to tell you what the player did wrong in general.
By contrast, the shortcomings of today's LLMs seem pretty obvious to me.
The (in)ability to recognize a strange move’s brilliance might depend on the complexity of the game. The real world is much more complex than any board game.
of course we can have AGI (damned if we don't) because we put so much, it better works
but the problem we cant do that right because its so expensive, AGI is not matter of if but when
but even then it always about the cost
they just need to "MCP" it to robot body and it works (also part of reason why OpenAI buys a robotic company)
The complexity of achieving those might result in the "Centaur Era", when humans+computers are superior to either alone, lasting longer than the Centaur chess era, which spanned only 1-2 decades before engines like Stockfish made humans superfluous.
However, in well-defined domains, like medical diagnostics, it seems reasoning models alone are already superior to primary care physicians, according to at least 6 studies.
Ref: When Doctors With A.I. Are Outperformed by A.I. Alone by Dr. Eric Topol https://substack.com/@erictopol/p-156304196
Medical diagnosis relies heavily on knowledge, pattern recognition, a bunch of heuristics, educated guesses, luck, etc. These are all things LLMs do very well. They don't need a high degree of accuracy, because humans are already doing this work with a pretty low degree of accuracy. They just have to be a little more accurate.
Alphazero also doesn't need training data as input--it's generated by game-play. The information fed in is just game rules. Theoretically should also be possible in research math. Less so in programming b/c we care about less rigid things like style. But if you rigorously defined the objective, training data should also be not necessary.
This is wrong, it wasn't just fed the rules, it was also fed a harness that did test viable moves and searched for optimal ones using a depth first search method.
Without that harness it would not have gained superhuman performance, such a harness is easy to make for Go but not as easy to make for more complex things. You will find the harder it is to make an effective such harness for a topic the harder it is to solve for AI models, it is relatively easy to make a good such harness for very well defined programming problems like competitive programming but much much harder for general purpose programming.
Then that is not a general algorithm and results from it doesn't apply to other problems.
If you mean symbolic reasoning, well it's pretty obvious that they aren't doing it since they fail basic arithmetic.
They can convincingly mimic human thought but the illusion falls flat at further inspection.
Calculators have been better than humans at arithmetic for well over half a century. Calculators can reason?
If that's your take-away from that paper, it seems you've arrived at the wrong conclusion. It's not that it's "fake", it's that it doesn't give the full picture, and if you only rely on CoT to catch "undesirable" behavior, you'll miss a lot. There is a lot more nuance than you allude to, from the paper itself:
> These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out.
A more "proper" approach would be to work with sets of hypotheses and to conduct tests to exclude alternative explanations gradually - which medics call "DD" (differential diagnosis). Sadly, this is often not systematically done, and instead people jump on the first diagnosis and try if the intervention "fixes" things.
So I agree there are huge gains from "low hanging fruits" to be expected in the medical domain.
AI producing visual art has only flooded the internet with "slop", the commonly accepted term. It's something that meets the bare criteria, but falls short in producing anything actually enjoyable or worth anyone's time.
However, even artists need supporting materials and tooling that meet bare criteria. Some care what kind of wood their brush is made from, but I'd guess most do not.
I suspect it'll prove useless at the heart of almost every art form, but powerful at the periphery.
But IRL? Lots of measures exist, from money to votes to exam scores, and a big part of the problem is Goodhart's law — that the easy-to-define measures aren't sufficiently good at capturing what we care about, so we must not optimise too hard for those scores.
Winning or losing a Go game is a much shorter term objective than making or losing money at a job.
> But IRL? Lots of measures exist
No, not that are shorter term than winning or losing a Go game. A game of Go is very short, much much shorter than the time it takes for a human to get fired for incompetence.
I agree the time horizon of current SOTA models isn't particularly impressive. Doesn't matter in this point.
Are you simply referring to games having a defined win/loss reward function?
Because pretty sure Alpha Go was ground breaking also because it was self taught, by playing itself, there were no training materials. Unless you say the rules of the game itself is the constraint.
But even then, from move to move, there are huge decisions to be made that are NOT easily defined with a win/loss reward function. Especially early game, there are many moves to make that don't obviously have an objective score to optimize against.
You could make the big leap and say that GO is so open ended, that it does model Life.
"artificial" maybe I should have said "synthetic"? I mean the computer can teach itself.
"constrained" the game has rules that can be evaluated
and as to the other -- I don't know what to tell you, I don't think anything I said is inconsistent with the below quotes.
It's clearly not just a generic LLM, and it's only possible to generate a billion training examples for it to play against itself because synthetic data is valid. And synthetic data contains training examples no human has ever done, which is why it's not at all surprising it did stuff humans never would try. A LLM would just try patterns that, at best, are published in human-generated go game histories or synthesized from them. I think this inherently limits the amount of exploration it can do of the game space, and similarly would be much less likely to generate novel moves.
https://en.wikipedia.org/wiki/AlphaGo
> As of 2016, AlphaGo's algorithm uses a combination of machine learning and tree search techniques, combined with extensive training, both from human and computer play. It uses Monte Carlo tree search, guided by a "value network" and a "policy network", both implemented using deep neural network technology.[5][4] A limited amount of game-specific feature detection pre-processing (for example, to highlight whether a move matches a nakade pattern) is applied to the input before it is sent to the neural networks.[4] The networks are convolutional neural networks with 12 layers, trained by reinforcement learning.[4]
> The system's neural networks were initially bootstrapped from human gameplay expertise. AlphaGo was initially trained to mimic human play by attempting to match the moves of expert players from recorded historical games, using a database of around 30 million moves.[21] Once it had reached a certain degree of proficiency, it was trained further by being set to play large numbers of games against other instances of itself, using reinforcement learning to improve its play.[5] To avoid "disrespectfully" wasting its opponent's time, the program is specifically programmed to resign if its assessment of win probability falls beneath a certain threshold; for the match against Lee, the resignation threshold was set to 20%.[64]
I was miss-remembering the order of how things happened.
AlphaZero, another iteration after the famous matches, was trained without human data.
"AlphaGo's team published an article in the journal Nature on 19 October 2017, introducing AlphaGo Zero, a version without human data and stronger than any previous human-champion-defeating version.[52] By playing games against itself, AlphaGo Zero surpassed the strength of AlphaGo Lee in three days by winning 100 games to 0, reached the level of AlphaGo Master in 21 days, and exceeded all the old versions in 40 days.[53]"
Unless it's a MUCH bigger play where through some butterfly effect it wants me to fail at something so I can succeed at something else.
My real name is John Connor by the way ;)
They are good at framing what is going on and going over general plans and walking through some calculations and potential tactics. But I wouldn't say even really strong players like Leko, Polgar, Anand will have greater insights in a Magnus-Fabi game without the engine.
I think large language models have the same future as supersonic jet travel. It’s usefulness will fail to realize, with traditional models being good enough but for a fraction of the price, while some startups keep trying to push this technology but meanwhile consumers keep rejecting it.
Unlike supersonic passenger jet travel, which is possible and happened, but never had much of an impact on the wider economy, because it never caught on.
That said, supersonic flight is yet very much a thing in military circles …
AI is a bit like railways in the 19th century: once you train the model (= once you put down the track), actually running the inference (= running your trains) is comparatively cheap.
Even if the companies later go bankrupt and investors lose interest, the trained models are still there (= the rails stay in place).
That was reasonably common in the US: some promising company would get British (and German etc) investors to put up money to lay down tracks. Later the American company would go bust, but the rails stayed in America.
The large language models are not that much better than a single artist / programmer / technical writer (in fact they are significantly worse) working for a couple of hours. Modern tools do indeed increase the productivity of workers to the extent where AI generated content is not worth it in most (all?) industries (unless you are very cheap; but then maybe your workers will organize against you).
If we want to keep the railway analogy, training an AI model in 2025 is like building a railway line in 2025 where there is already a highway, and the highway is already sufficient for the traffic it gets, and won’t require expansion in the foreseeable future.
That's like saying sitting on the train for an hour isn't better than walking for a day?
> [...] (unless you are very cheap; but then maybe your workers will organize against you).
I don't understand that. Did workers organise against vacuum cleaners? And what do eg new companies care about organised workers, if they don't hire them in the first place?
Dock workers organised against container shipping. They mostly succeeded in old established ports being sidelined in favour of newer, less annoying ports.
No, that’s not it at all. Hiring a qualified worker for a few hours—or having one on staff is not like walking for a day vs. riding a train. First of all, the train is capable of carrying a ton of cargo which you will never be able to on foot, unless you have some horses or mules with you. So having a train line offers you capabilities that simply didn’t exist before (unless you had a canal or a navigable river that goes to your destination). LLMs offers no new capabilities. The content it generates is precisely the same (except its worse) as the content a qualified worker can give you in a couple of hours.
Another difference is that most content can wait the couple of hours it takes the skilled worker to create it, the products you can deliver via train may spoil if carried on foot (even if carried by a horse). A farmer can go back tending the crops after having dropped the cargo at the station, but will be absent for a couple of days if they need to carry it on foot. etc. etc. None of these is applicable for generated content.
> Did workers organize against vacuum cleaners?
Workers have already organized (and won) against generative AI. https://en.wikipedia.org/wiki/2023_Writers_Guild_of_America_...
> Dock workers organised against container shipping. They mostly succeeded in old established ports being sidelined in favour of newer, less annoying ports.
I think you are talking about the 1971 ILWU strike. https://www.ilwu.org/history/the-ilwu-story/
But this is not true. Dock workers didn’t organized against mechanization and automation of ports, they organized against mass layoffs and dangerous working conditions as ports got more automated. Port companies would use the automation as an excuse to engage in mass layoffs, leaving far too few workers tending far to much cargo over far to many hours. This resulted in fatigued workers making mistakes which often resulted in serious injuries and even deaths. The 2022 US railroad strike was for precisely the same reason.
I wouldn't just willy nilly turn my daughter's drawings into cartoons, if I had to bother a trained professional about it.
A few hours of a qualified worker's time takes a couple hundred bucks at minimum. And it takes at least a couple of hours to turn around the task.
Your argument seems a bit like web search being useless, because we have highly trained librarians.
Similar for electronic computers vs human computers.
> I think you are talking about the 1971 ILWU strike. https://www.ilwu.org/history/the-ilwu-story/
No, not really. I have a more global view in mind, eg Felixtowe vs London.
And, yes, you do mechanisation so that you can save on labour. Mass layoffs are just one expression of this (when you don't have enough natural attrition from people quitting).
You seem very keen on the American labour movements? There's another interesting thing to learn from history here: industry will move elsewhere, when labour movements get too annoying. Both to other parts of the country, and to other parts of the world.
Even the fancy models where you need to buy compute (rails) that's about the price of a new car, they have a power draw of ~700W[0] while running inference at 50 tokens/second.
But!
The constraint with current hardware isn't compute, the models are mostly constrained by RAM bandwidth: back of the envelope estimate says that e.g. if Apple took the compute already in their iPhones and reengineered the chips to have 256 GB of RAM and sufficient bandwidth to not be constrained by it, models that size could run locally for a few minutes before hitting thermal limits (because it's a phone), but we're still only talking one-or-two-digit watts.
[0] https://resources.nvidia.com/en-us-gpu-resources/hpc-datashe...
[1] Testing of Mistral Large, a 123-billion parameter model, on a cluster of 8xH200 getting just over 400 tokens/second, so per 700W device one gets 400/8=50 tokens/second: https://www.baseten.co/blog/evaluating-nvidia-h200-gpus-for-...
That hardware cost Apple tens of billions to develop and what you're talking about in term of "just the hardware needed" is so far beyond consumer hardware it's funny. Fairly sure most Windows laptops are still sold with 8GB RAM and basically 512MB of VRAM (probably less), practically the same thing for Android phones.
I was thinking of building a local LLM powered search engine but basically nobody outside of a handful of techies would be able to run it + their regular software.
Despite which, they sell them as consumer devices.
> and what you're talking about in term of "just the hardware needed" is so far beyond consumer hardware it's funny.
Not as big a gap as you might expect. M4 chip (as used in iPads) has "28 billion transistors built using a second-generation 3-nanometer technology" - https://www.apple.com/newsroom/2024/05/apple-introduces-m4-c...
Apple don't sell M4 chips separately, but the general best-guess I've seen seems to be they're in the $120 range as a cost to Apple. Certainly it can't exceed the list price of the cheapest Mac mini with one (US$599).
As bleeding-edge tech, those are expensive transistors, but still 10 of them would have enough transistors for 256 GB of RAM plus all the compute each chip already has. Actual RAM is much cheaper than that.
10x the price of the cheapest Mac Mini is $6k… but you could then save $400 by getting a Mac Studio with 256 GB RAM. The max power consumption (of that desktop computer but with double that, 512 GB RAM) is 270 W, representing an absolute upper bound: if you're doing inference you're probably using a fraction of the compute, because inference is RAM limited not compute limited.
This is also very close to the same price as this phone, which I think is a silly phone, but it's a phone and it exists and it's this price and that's all that matters: https://www.amazon.com/VERTU-IRONFLIP-Unlocked-Smartphone-Fo...
But irregardless, I'd like to emphasise that these chips aren't even trying to be good at LLMs. Not even Apple's Neural Engine is really trying to do that, NPUs (like the Neural Engine) are all focused on what AI looked like it was going to be several years back, not what current models are actually like today. (And given how fast this moves, it's not even clear to me that they were wrong or that they should be optimised for what current models look like today).
> Fairly sure most Windows laptops are still sold with 8GB RAM and basically 512MB of VRAM (probably less), practically the same thing for Android phones.
That sounds exceptionally low even for budget laptops. Only examples I can find are the sub-€300 budget range and refurbished devices.
For phones, there is currently very little market for this in phones, the limit is not because it's an inconceivable challenge. Same deal as thermal imaging cameras in this regard.
> I was thinking of building a local LLM powered search engine but basically nobody outside of a handful of techies would be able to run it + their regular software.
This has been a standard database tool for a while already. Vector databases, RAG, etc.
Oh, please show me the consumer version of this. I'll wait. I want to point and click.
Similar story for the consumer devices with cheap unified 256GB of RAM.
But again the original argument was that they can run forever because inference is cheap, not cheap enough if you’re losing money on it.
Air travel of course taking over is the main reason for all of this but the costs sunk into the rails are lost or ROI curtailed by market force and obsolescence.
Very few people want to invest more: the private sector doesn't want to because they'll never see the return, the governments don't want to because the returns are spread over their great-great-grandchildren's lives and that doesn't get them re-elected in the next n<=5 (because this isn't just a USA problem) years.
Even the German government dragged its feet over rail investment, but they're finally embarrassed enough by the network problems to invest in all the things.
Remember these outsourcing firms that essentially only offer warm bodies that speak English? They are certainly already feeling the impact. (And we see that in labour market statistics for eg the Philippines, where this is/was a big business.)
And this is just one example. You could ask your favourite LLM about a rundown of the major impacts we can already see.
There's no emotional warmth involved in manning a call centre and explicitly being confined to a script and having no power to make your own decisions to help the customer.
'Warm body' is just a term that has nothing to do with emotional warmth. I might just as well have called them 'body shops', even though it's of no consequence that the people involved have actual bodies.
> A frigging robot solving your unsolvable problem ? You can try, but witness the backlash.
Front line call centre workers aren't solving your unsolvable problems, either. Just the opposite.
And why are you talking in the hypothetical? The impact on call centres etc is already visible in the statistics.
And with trains people paid for a ticket and a hard good “travel”
Ai so far gives you what?
Demand for AI is insanely high. They can't make chips fast enough to meet customer demand. The energy industry is transforming to try to meet the demand.
Whomever is telling you that consumers are rejecting it is lying to you, and you should honestly probably reevaluate where you get your information. Because it's not serving you well.
No, instead it'll be the new calculator that you can use to lazy-draft an email on your 1.5 hour Ryanair economy flight to the South. Both unthinkable luxuries just decades ago, but neither of which have transformed humanity profoundly.
Currently market data is showing a very high demand for AI.
These arguments come down to "thumbs down to AI". If people just said that it would at least be an honest argument. But pretending that consumers don't want LLMs when they're some of the most popular apps in the history of mankind is not a defensible position
Reasons for market data seemingly showing high demand without there actually being one include: Market manipulation (including marketing campaigns), artificial or inflated demand, forced usage, hype, etc. As an example NFTs, Bitcoin, and supersonic jet travel all had “an insane market data” which seemed at the time to show that there was a huge demand for these things.
My prediction is that we are in the early Concord era of supersonic jet travel and Boeing is racing to catch up to the promise of this technology. Except that in an unregulated market such as the current tech market, we have forgone all the safety and security measures and the Concord has made its first passenger flight in 1969 (as opposed to 1976), with tons of fan fare and all flights fully booked months in advance.
Note that in the 1960 it was market forecasts had the demand for Concord to build 350 airplanes by 1980, and at the time the first prototypes were flying they had 74 options. Only 20 were every built for passenger flight.
But! We here are not typical callers necessarily. How many IT calls for general population can be served efficiently (for both parties) with a quality chatbot?
And lest we think I'm being elitist - let's take an area I am not proficient in - such as HR, where I am "general population".
Our internal corporate chatbot has turned from "atrocious insult to man and God's" 7 years ago, to "far more efficiently than friendly but underpaid and inexperienced human being 3 countries away answering my incessant questions of what holidays do I have again, how many sick days do I have and how do I enter them, how do I process retirement, how do I enter my expenses, what's the difference between short and long term disability" etc etc. And it has a button for "start a complex hr case / engage a human being" for edge cases,so internally it works very well.
This is a narrow anecdata about notion of service support chatbot, don't infere (hah) any further claims about morality, economy or future of LLMs.
Woah there cowboy, slow down a little.
Demand for chips is come from the inference providers. Demand for inference was (and still is) being sold at below cost. OpenAI, for example, has a spend rate of $5b per month on revenues of $0.5b per month.
They are literally selling a dollar for actual 10c. Of course "demand" is going to be high.
This is definitely wrong, last year it was $725m/month expenses and $300m/month revenue. Looks like the nearly-2:1 ratio is also expected for this year: https://taptwicedigital.com/stats/openai
This also includes the cost of training new models, so I'm still not at all sure if inference is sold at-cost or not.
(To be clear, I'm not criticising the person I'm replying to.)
I tend to rough-estimate it based on known compute/electricity costs for open weights models etc., but what evidence I do have is loose enough that I'm willing to believe a factor of 2 per standard deviation of probability in either direction at the moment, so long as someone comes with receipts.
Subscription revenue and corresponding service provision are also a big question, because those will almost always be either under- or over-used, never precisely balanced.
It looks like you're using "expenses" to mean "opex". I said "spend rate", because they're spending that money (i.e. the sum of both opex and capex). The reason I include the capex is because their projections towards profitability, as stated by them many times, is based on getting the compute online. They don't claim any sort of profitability without that capex (and even with that capex, it's a little bit iffy)
This includes the Stargate project (they're committed for $10b - $20b (reports vary) before the end of 2025), they've paid roughly $10b to Microsoft for compute for 2025. Oracle is (or already has) committed $40b in GPUs for Stargate and Softbank has committments to Stargate independently of OpenAI.
> Looks like the nearly-2:1 ratio is also expected for this year: https://taptwicedigital.com/stats/openai
I find it hard to trust these numbers[1]: The $40b funding was not in cash right now, and depends on Softbank for $30b with Softbank syndicating the remaining $10b. Softbank themselves don't have cash of $30b and has to get a loan to reach that amount. Softbank did provide $7.5b in cash, with milestones for the remainder. That was in May 2025. In August that money had run out and OpenAI did another raise of $8.3b.
In short, in the last two to three months, OpenAI spent $5b/month on revenues of $0.5b/m. They are also depending on Softbank coming through with the rest of the $40b before end of 2025 ($30b in cash and $10b by syndicating other investors into it) because their commitments require that extra cash.
Come Jan-2026, OpenAI would have received, and spent most of, $60b for 2025, with a projected revenue $12b-$13b.
---------------------------------
[1] Now, true, we are all going off rumours here (as this is not a public company, we don't have any visibility into the actual numbers), but some numbers match up with what public info there is and some don't.
I took their losses and added it to their revenue. That seems like that sum would equal expenses.
> The $40b funding was not in cash right now,
Does this matter? I'm not counting it as revenue.
> In short, in the last two to three months, OpenAI spent $5b/month on revenues of $0.5b/m.
You're repeating the same claim as before, I've not seen any evidence to support your numbers.
The evidence I linked you to suggests the 2025 average will be double that revenue, $1bn/month, at an expense of ($9bn loss after $12bn revenue / 12 months = $21bn / 12 months) = $1.75bn/month
> Does this matter? I'm not counting it as revenue.
Well, yes, because they forecast spending all of it by end of 2025, and they moved up their last round ($8.3b) by a month or two because they needed the money.
My point was, they received a cash injection of $10b (first part of the $40b raise) and that lasted only two months.
>> In short, in the last two to three months, OpenAI spent $5b/month on revenues of $0.5b/m.
> You're repeating the same claim as before, I've not seen any evidence to support your numbers.
Briefly, we don't really have visibility into their numbers. What we do have visibility into is how much cash they needed between two points (Specifically, the months of June and July). We also know what their spending commitment is (to their capex suppliers) for 2025. That's what I'm using.
They had $10b injected at the start of June. They needed $8.3b at the end of July.
Chatgpt, claude, gemini in chatbot or coding agent form? Great stuff, saves me some googling.
The same AI popping up in an e-mail, chat or spreadsheet tool? No thanks, normal people don't need an AI summary of a 200 word e-mail or slack thread. And if I've paid a guy a month's salary to write a report on something, of course I'll find 30 minutes to read it cover-to-cover.