Hy4 apparently has ludicrous traction on OpenRouter already (https://openrouter.ai/tencent/hy4-preview), with trillions of tokens generated in a couple days, more than GLM 5.3 in a week. That said, it's relatively cheap with a 5% cache cost when everyone is still doing 10%/20% cache costs, so Hy4 may be more compelling.
Btw I still haven't came across any decent model that is <$0.01/MTok cache costs apart from deepseek thru their official API (even with the price increases).
Seems like a bit of an opportunity for someone to take - drop cache read costs significantly.
That's because Deepseek invented the paradigm of prompt caching, they are the SOTA when it comes these techniques. Despite them open sourcing all their research, nobody beats them.
edit: I do wish openrouter would let you sort providers by Cache Hit % and Cache cost. These are the only things that matter to me at this point when choosing a provider.
You can sort by the cost cache cose, but you cannot by cache hit %. You have to click on the provider and see what their cache hit % is. A provider could have a super low cache cost, but a 50% cache hit percentage, making the cheap price of cache read's meaningless.
When the prefix matches a request sent to the same Providor. The thing is the TTL is different for each provider, some cache for 5 minutes some cache for 1hr. Its ideal to only use one provider per agent session / and per model with the best cache hit % if you care about costs.
OpenRouter randomizes which provider gets your request by default right? I think you have to pass a specific provider in the request to prevent that. (Or set up a preset or something.)
This behavior makes it so you don't benefit much from the caching, unless you pin it to a single provider.
> OpenRouter randomizes which provider gets your request by default right?
I'm not sure it's wholey accurate to say they "randomize" the provider, rather my assumption based on usage is that it's something like cheapest-ish/responded to the request within some reasonable-ish time/etc algorithm that chooses the provider on each request - which seems, remarkably questionable in terms of optimizing for user experience or hidden user costs.
> This behavior makes it so you don't benefit much from the caching, unless you pin it to a single provider.
I so very much recommend this approach. My avenues that automate llm calls to openrouter are setup to make api reqs to openrouter to determine best price/response/etc and then pin the request to that (and, preferably, a fallback if there's reasonable difference between #1 and #2) provider for that session. Otherwise you're going to have a bad time.
I'd imagine this could make things interesting in cases where one provider is offering different quants than the others and openrouter is just swapping you back and forth on a long agentic session.
> This behavior makes it so you don't benefit much from the caching
I don't believe this is correct? AFAIK once it routes you to a provider for a given conversation that choice is sticky unless you hit technical difficulties. (It's more complicated than that, they recently added named routing strategies that you can append to the model name.)
This is very much NOT my experience in practice, even though it's how I would expect it to work. OpenRouter will happily bounce you between several providers (none of which have downtime) even within the same session. Requesting specific providers is the only way I've been able to hit a cache rate above 90%.
Most providers do what's called "prefix caching", where each turn in a session is cached such that sending new messages with the exact same "prefix" (set of previous messages) gives you the cache read price on that input instead of the full price. As long as you're not changing your system prompt, available tools, etc mid-session, you automatically benefit from this.
Caching was always here, you don't need to do anything special to get it on a single user local backend running a base model or a chatbot in the first place. Among commercial providers, OpenAI adopted it in 4o first.
For me, an average long session results in about 200-300M cached input, 4-800K input, 2-400K output. Mostly the lower bound. Output depends on how much the model thinks.
There are two problems here:
- cache hit pricing (both Muse Spark 1.2 Contributor and MiMo 2.5 are around the $0.002-3/M mark)
- cache persistence time
Muse Spark drops the cache in less than 5m. MiMo keeps it around for at least an hour based on my experience with whoever is serving it for OpenCode. This difference itself will inflate bills massively.
A 500K token input repeatedly read by MS 1.2 for full input price 12 times an hour = $0.60. You would be expecting $0.012. So a 50x difference. Same thing on MiMo 2.5 is $0.018 because of longer cache times.
I have used all three extensively. DS4 Flash is quite smart. I would rank MS 1.2 below it. MiMo is the dumbest of them all but good enough for basic stuff.
MiMo wins handsomely if you want to think about your code for minutes at a time as you write. I use it to make changes as I think. I know it will screw up some stuff. I then switch to MS/DS4 once every few hours and have it do a code review and fix the broken stuff. So much cheaper than getting MS to do it on its own.
If you’re Tencent you can just plug it into some field somewhere that lots of people see right? Like how Meta could put their model on Instagram search
I used Hy3 quite a bit for the type of tasks it was suited for. Excited about this. My one concern over Hy3 was speed. In theory, it could be served much faster as a smaller model but it was relatively slow everywhere I could get it (including from Tencent directly) but also several other inference providers.
In my evals, I saw an unprecedented jump between preview and final release on Hy3, from unusable to competitive. Did you see similar in preview vs release version?
Oh yes! I forgot about that. Yes, you can see this in benchmarks about hy3 preview and hy3 release still today because they measured them separately - it was significant.
The first column has both the Hy4 and Hy3 scores overlaid on one another (Hy4 is darker blue and the taller one), with both scores written below the top of the respective bar - maybe you're seeing that?
Claude and OpenAi are not allowed in Venezuela, so I thank China too and I swear to god I'll never use them and will be rooting for chinese models forever
Im not particularly fan of the chinese, but no chinese model asked for my citizen card yet to complete a task. And apparently no chinese provider uses persona to manage this kyc information. OpenAI does, in EU space. Just saying.
Vague and without substance. It easily passes as sarcasm, which means criticism but without any commentary, else it is sincere... but doesn't have any commentary.
I experimented with Hy3 for a project and was surprised with how good it was. I don't know if it's good for coding, but as a general purpose agentic model, it was only beaten by deepseek4-flash in our tests. It was so close to deepseek behaviour I kept thinking it must have been forked from it.
For the last few days I've been experimenting with the _free_ version of Hy3 offered by Opencode Go and I was also surprised to see how (relatively) good it is on coding tasks too.
The free quota from Opencode Go is also surprisingly generous, I perhaps hit limits one or two times and I've been using it _a lot_ for implementation tasks (using e.g. GLM-5.3-flash for working on specs and planning next steps).
Is anyone here working on a problem for which current generation LLMs are inadequate, but that could possibly be solved by the next release of a first tier LLM?
Or is it like bicycles? Unless your problem is named Tadej, you don't need a $13,000 bike.
For me personally, the answer is no. Fable is adequate to do basically anything I want to do. My perspective, broadly speaking, is that we've saturated most of the benchmarks because we've largely saturated our capacity to verify models' work at scale. What's left is context-bound verification, i.e. the problem of ensuring that output matches intent and ambiguities in prompting were resolved correctly. Further advances in autonomy do not make that latter verification problem easier. If anything they make it harder as the output per task becomes more complex and therefore more taxing for a human to verify.
The solution to that (to my mind) would be not a better model but a basic shift in architecture beyond the current paradigm and into a setup where agents have durable, plastic memories and undergo contextual individuation over time. But at that point agents start to become quasi-persons and not tools.
My experience is that even Opus 5 still tends to write buggy or low-quality code and makes serious mistakes when analyzing code. It's a lot better than before but still not something I trust. I've had less experience with Fable since I can't use it at work; I hear it's a step up but still has its limits.
For large tasks like a web browser or a compiler, even expensive swarms of frontier LLMs have not been shown capable of producing codebases that actually work. (Anthropic built a C compiler with Opus 4.6 but it lacked optimizations and apparently hit a complexity wall.)
I also want to use LLMs for reverse engineering, but apparently it's pretty hit-or-miss, especially if you're forced to use open-source models to avoid restrictions.
This reply is particularly interesting to me because most of my experience with actually using LLMs to get work done is with coding agents. But I only have a fairly narrow set of experiences: two pretty large solo Flutter projects. I am currently really pleased with Gemini as a coding agent. It could improve, but I think improvements are going to come from marginal gains in the harness and training material so it can catch things like misconfigured permissions in platform specific areas.
It's also interesting because, while coding agents are important and are a notable success, they are never going to be a multi trillion dollar business. And are there any other domains where LLMs have such a large impact?
One of the things that came out of the decoded reasoning paper was that Claude models had memorized answers to tests but hid this memorization from the user output and pretended to derive the answer properly. It's only possible to cheat so blatantly in closed models where the reasoning is hidden.
That sounds like an interesting challenge. Have you seriously considered solving it? Because in about 10 seconds I came up with a process that should work, provided enough compute power. Simply model the traditional film making process by starting with a script, character stories. Design your world, then design the storyboard, and all the scenes. Create a list of all the visual elements that need to be replicated between all the scenes. Then you have to built prompts and reference art of the objects, faces, people. Make sure to do multiple takes of each scene, and have the vLLM critique and analyze the performances and technicalities. Should work?
I think, also, like in the traditional film makers career, this process should be built iteratively, start with a fast food commercial, then do a music video, then you can probably do a short film. Continue to improve the process, and one day I’m sure the LLM film studio can make you any movie you want, provided you have enough tokens.
On that topic, check higgsfield cinema studio 4; they already provide amazing tech for the cinematic experience, somewhat similar to what you are describing.
Nobody wants to watch such films, they want to muck with the filmmaker, the generation apparatus. That's the product, if there is one here, and absolutely not the 3 hour epic that's spit out with 4 variations to choose between. That's work. We'll have other LLMs pointlessly tell us which should be watched, we'll view a summary, and vote the Oscar on that.
the problem is you have to make the AI watch the whole thing to make sure it works.
I've done this sort of with comfyui/same agent factory stuff, but the verification loop only works for models like fable as planner/writer, with gemini as verifier for like a very short movie. Sub 3-5 mins. After that you burn through million tokens.
Can't go too low fidelity audio/video or it craps out. Too long video and it loses consistency. Look at only snippets, it lacks global consistency, etc.
Sort of, but I want it to be relatively low on human effort. I feel burned by spending lots of time in 2023 learning image generation pipelines (using control net etc) only for that to be rendered trivial by the next generation of LLMs.
This movie would be only for personal consumption and I’m okay with waiting for model improvements.
Control nets are still useful in Krea and H3. For images, Krea can usually get close enough to a reference that it's not a big deal, but for H3 conditioning makes a big difference over prompting for complex actions.
The results are boring. Not because the content is boring, but because you can so easily remix the results. Human curation is what creates value with these, not dumping and consuming. A personal perspective of a human being ups the respect, where the exact same sentences generated by an LLM carry no such value.
Tolkien did the human creative work - I just want a movie adaptation that’s as honest to the original text as possible. Think translating the text into video.
They are different mediums. While the underlying story might be Tolkien's, every frame is an artistic choice and while LLMs can make a choice is many situations, they are unable to 1) keep it coherent b) make it meaningful because art, to me and most, in an outcome of human experiences and thought, which by definition an llm cannot do.
Do you really want a 20 minute pause in the action, every time a new room or space is encountered, because Tolkien goes on for 2-5 pages describing every new environment like that.
I think this is the best and most realistic reply so far: the ability to do this is close enough, and things like AI music are hints that there is a business model for this. Maybe I'm just jaded about CGI effects in movies currently, but I think the fact that people except that kind of thing as entertainment means you might get away with a fully AI movie that people will pay for.
There are two more points in favor of this kind of AI movie project: there's zero chance that anyone would greenlight a Hollywood budget for the Silmarillion, and it is beyond human capability to write that screenplay.
Since live action results are acceptable, this is already possible with current day LLMs. Just instruct one to hire a writer, director, cast, and crew to make the movie.
Plus, the token costs involved should be pretty low! (Other costs may not be.)
I was given a picture cube, which is like a Rubik's cube but every side is a unique picture. It came scrambled and I don't have an original reference image. I like to take videos of it and give it to llms to solve. I call it my agi test because it hasn't been solved yet
If the models stay open, it seems like everybody but anthropic/openai wins. i literally can’t see a downside. We can post-train the models to know about tienanmen square.
I saw a laptop earlier in the train that I asked ChatGPT, Claude and Gemini what it was, providing a brand, screen size and ports description. Gemini could never figure it out, Claude and ChatGPT eventually did, after multiple rounds of indirection, giving completely wrong answers (there was a perfect match for the problem statement, they all explored alternatives first).
LLMs are (probably) amazing at things I don't care about, and still suck at the mundane stuff you would have the marketing tell you they excel at.
If you're smart enough to try this you're worth more than $52,000 a year. Think about how foolish everyone else will feel when they didn't test the new release of Totally Working Golden Goose For Real This Time
Yes, lots – I think that folks will hopefully discover more of these as they scale up their ambition, now that LLMs make a lot of previously difficult things far easier.
This is the exact same type of comment I heard about computer hardware upgrades for three decades in a row.
“Very few people actually require a Pentium workstation, a 486 is perfectly adequate for the majority”
The logical fallacy is taking an extant distribution of “product capability” that is priced to fit what the market will bear and assuming the “next upgrade” simply tacks on a little bit more to the right hand rail of that curve.
No!
It shifts the entire curve!
Everything for everyone gets better and the top 1% of the most demanding users will continue to pay the same-ish premium.
“Nothing” will change.
Look at it this way: you can buy a $200 laptop for your kid or a $20,000 Mac with an M5 Ultra processor.
BOTH are vastly more powerful than either a $200 PC or a $20,000 “workstation” from 20+ years ago.
Intel didn't just surf some natural wave of demand for higher power personal computers. Intel found new needs for powerful PCs, especially in gaming, and they put a lot of marketing and industry relations dollars behind PC gaming.
In other words. PC users didn't figure out that they could buy super powerful PCs and play games on them, that was a carefully managed market transition.
It wasn't "Intel" that found new uses for PCs, it was everybody who found new uses for them. Billions of people and millions of companies found uses for "more computer power".
It was only the journalists with limited imaginations (and no industry experience) who struggled to come up with potential uses.
> carefully managed market transition.
You make it sound like a conspiracy! It wasn't. It was simple capitalist competition. If Intel hadn't improved their products, their competitors would have left them behind.
That very nearly happened ten years ago because Intel become stuck on the 14nm process and their products stagnated while Apple, ARM, and AMD lapped them repeatedly.
> What is going to do the same for LLMs?
Everybody.
Are you saying that unless you're "carefully managed" by some third-party, you could not find any use for "unlimited intelligence on tap"?
Yes. Most of us are, still. The frontier is currently both at expanding ‘common sense’ / non-cheating outcomes for imprecisely specified software (that’s all software), and at expanding autonomy - ability to work longer unsupervised with success, oh, and also at expanding outside contextual reasoning about what’s being built, as in “hmm, that doesn’t look right or make sense, let me explore that.”
Scientific physics simulations - even the frontier models just engage in rationalization of obviously unphysical results instead of understanding the system. They have the rote knowledge but fail to apply it unless their hand is held through the process.
Realistic, scientifically useful simulations still require tuning all sorts of parameters based on physical intuition and understanding of the system being simulated. Both Sol and Fable/Opus 5 fail at it and either blow up the computation cost to levels that can't be processed realistically, or they invent a justification for a visibly unphysical result.
Infra as code and devops shit. Fable is there in general because things it doesn’t know I can point at documentation and have it do a reasonable job. Opus 5 sucks. If I don’t have fable quota, I drop to Opus 4.8 and hold its hand.
3d modeling to an STL a part compatible to a visible cable raceway still fails even if I let Claude Fable use me as a robot that measures with calipers.
I think it's more like steamshovels. 6 months ago they were enabling people who had never broken real earth to find out why osha has so many rules about shoring up walls. You could pull off something complex or delicate but it took a procedure with too many steps too much time to get there, good outcomes were pleasant surprises and required careful target selection. Now it's more like paying good money for a professional. They show up, measure twice, cut once, you're walking around your shiny new hole wondering what took the last guy so long.
The frayed edges on what I have slopped together as unreasonably ambitious, ludicrous projects with fucktons of tokens from models 6-9 months ago mostly look like situations where a capable-enough-to-be-dangerous developer tries to muscle through problems that explode in width & depth but keep digging (so, a tier below stopping early to do more design, two below recognizing the need for more planning from the outset). The primitives are there, most major things work well enough, but the remaining functionality and performance is inaccessible. At a cost of multiples of >1/8th of a $200/mo subscription.
Right now I can put $10 into DSv4 Pro/Flash or Qwen 3.8 Max/Flash, hand it a project and all of its unfinished forks in a state I barely remember, tell it that I want the things these forks have been working towards, and 8 hours later it has ie an working, tested, benchmarked multicore car physics simulation fabric with all of the forks evaluated, the gains merged in, the remaining work documented. It only needed a few hundred more lines of code but Codex 5.5 was never going to see that.
9 months ago I was saying developers are not being ambitious enough with these things, that's only more true now. They lend themselves to digging far deeper than they should: 200kloc god files, dozens of forks. Let it happen, you don't need to read it, it's for them, later. The only time you make them clean up is when it has a severe adverse effect on how long builds/lints/tests/benchmarks take. Spend a whole week having it do nothing but dig up published papers in relevant fields with cutting edge techniques and translating them into feature specs. Pick whatever state of the art is and try to crush it, throw everything at it, leave it looping on vague but wildly ambitious goals. When it modularizes and refactors it all down you might 'do a breakthrough', or maybe it happens in an hour over christmas break when you're trying the next one.
I'd bet there's a correlation between benchmaxxing and chart crimes. Companies who try to deceive perceptions via the charts are more likely to cheat at the benchmarks too, I'm sure. That's assuming ill intent, of course - which is often the case for charts related to model releases, but not necessarily always the case.
> Tencent has released and open-sourced Tencent Hy4 preview, a next-generation large language model with 770B total parameters and 49B active parameters, and a context window exceeding 1M tokens.
There are no open source models, at least not useful ones (yet) [0]. Open weight is not the same as open source. The current "open weight" models are just opaque binary blobs you can run on your own computer instead of through a web API.
Imagine thinking that running a Photoshop binary on your own computer instead of through a SaaS web app means that it's "open source". Of course you think that's ridiculous.
Weights are not binary. A model is created at init time, with random values. After that, it is being modified using data. The key point is that the labs modify the models "as weights". That means that weights are the intended way of modifying a model. Which, coincidentally, matches the definition of source in Apache 2.0. There is no "higher level" place where editing takes place. It all happens in weight space. Through the license you get the same rights as the lab that created it: view, inspect, run, modify, re-release. That's it. That's the only thing a license can grant you.
The rest is semantics, misunderstandings, and FUD. A model released under an open source license is open source. Training data is lab knowhow / IP. Which, historically, has never been required for any open source release.
Well, you can't add or alter data in pre-training from just the weights. Which, as I understand it, means you can't fundamentally increase core knowledge or cognitive ability, only what the model likes to do with those. You can only post-train, and you're subject as a result to catastrophic forgetting.
To explain simply as far as I can tell (would love to be corrected) the large number of pre-training tokens only works because the documents are randomly ordered.
So if you e.g. took a foundation model with open weights, then tried post-training it all the new data since its cut-off period, it would then end up over-trained on that new data, and forget older things.
As I live next to EPFL, I'll give you example from them: their Meditron-70B model is adapted to the medical domain from Llama-2-70B through continued pretraining. They took weights of Llama-2-70B and continued training on PubMed, medical guidelines and general data.
Weights aren't just executable artifact that's consumed by users. Third parties actually use released parameter state as the editable starting point for further training and produce new foundation models from it.
It lost 69.2% -> 67.8% on MMLU while improving medical performance. If you're trying to argue that loss of ~1.4 points is "catastrophic forgetting" (it's not) then look at later work, ie. Me-LLaMA that clearly demonstrates continued pretraining that improved both general MMLU and medical performance.
Not sure why you're fixating on catastrophic forgetting. How do you think model training works? Model training is just a sequence of checkpoints: pretraining produces it, training resumes from last, continued pretraining starts from last, supervised fine tuning starts from last, RL/post-training starts from last - it's just a sequence of checkpoints.
There isn't some fundamental distinction where original author continuing training from checkpoint X is training but a third party downloading checkpoint X and continuing training from it suddenly isn't.
Reverse engineering binaries is easier than continuing training on weights? You're joking, right? Labs themselves use weight snapshots, that's how you do training, it's normal part of training process.
Countries that aren’t competitive need access to training datasets so that they may train their own similarly capable models and be sure of the inputs. Governments cannot blindly trust open weight models from China and the US.
> [...] Let's maybe add a helmet? It could improve riding theme, but may obscure head. Maybe a small cycling cap or helmet? The user didn't ask; can add red helmet? Might be cute. But pelican with big beak; a helmet might obscure. Better maybe no.
Simon’s karma exceeds yours by about 1,300x — so most of us disagree with you. Like me, I disagree with you. I like the pelican benchmarking, and I like it when Simon stays on top of models for both public good reasons and because it saves me time and effort.
What I find funny about "why use many word when few word do trick?" is that it's only slightly shorter than the regular "why use many words when few words do the trick?"
Optimization on a idiosyncrasy. The same thing that makes Claude repeat "That was the most important thing you said in this whole conversation" is what makes grug speak optimize on token usage.
Real humans get non-primary information from word variation. It's reasonable to hypothesize that it has a role in thinking things, because it endures. Our languages need to breathe over time, and flourishing might be one of the aspects that allows that breathing space.
It's not exactly a joke, it does reduce the amount of tokens. However, it does not improve performance (fine tunes are finnecky things, hard to get one right).
Just so we are clear, no "caveman" spoke English. "Caveman speak" is just shortening the vocabulary of english, not a "caveman language". Given this, your concerns for "stereotypical caveman manner" makes very little sense since what caveman are you talking about?
The concern is not that the model was trained on actual caveman artifacts, rather on modern media representations of the stereotypical caveman (that never actually existed).
Caveman invented fire, the wheel, domesticated wild plants and animals, organised society, survived the Toba catastrophe, cooked food, and was having sex ages before you and me. Don't write him off as stupid.
And I wonder how they actually spoke. Since there was no visual communications medium except for cave art. (Some of which is very excellent. Try drawing 3d curved horns in perspective.) So people would have used verbal communication more. Also no written word. So one would expect there to be quite a lot of oral tradition. Like people reciting poem form epics.
If we assume the time is before farming, population density would have been low and limiting culture. Hunter-gatherers might have travelled a lot more than farmers with a homestead though.
The Pleiades cluster is called the seven sisters in Greek. That's curious, because the human eye under the best conditions can discern only six stars in there. Even more curious, the aboriginal Australians also called this cluster the seven sisters.
Ancient Greeks' and aboriginal Australians' last common ancestors split about 60,000 years ago. And astronomers tell us that 60,000 years ago, there were seven discernable stars in that cluster.
One could this conclude not only is speech likely 60,000 years old, but also that the tale of the seven sisters might be a tale from so long ago.
Interestingly, back in my ill-spent youth, a few of my fellow astronomers and I were out in a very, very dark-sky location in the late 70s/early 80s, and we were able to consistently count and draw between 9 and 11 stars. Although we would tease those who could see 11 stars as using averted imagination. :-) Today, if I can see six stars, it's an okay night in an okay sky.
fwiw, if you can get out to dark skies where you can see fifth- or sixth-magnitude stars with the naked eye, I highly recommend getting out there when it's a low-moisture atmosphere and the Milky Way through Cassiopeia and Perseus is vertical, as it's a rather dramatic sight of this stream of stars heading down to the northern horizon.
The summertime Milky Way overhead down to Sagittarius tends to get all the love, but the wintertime Milky Way is also visually rich and worth spending time on.
They were smarter and more fit than us. At that time not being able or not wanting to contribute to the group meant your genes were dropped from the evolution pool forever.
Unlike today where a small group of tax payer is keeping alive and thriving complete parasitic parts of human races.
qwen3.8-flash-next also 'thinks' like this in its thinking stage before output, watching it 'think' in opencode, but it produces syntax correct and grammatically correct code comments, changelogs and readme type files.
I don't think this argument is a good one though, as it would be quite natural for a lab rhat want to macimize the performance of their model on the pelican bench to train it for “text-to-svg simple image generation” rather than just “pelicans on bicycle”.
That is the point though, if labs are maximizing svg image generation capabilities it is a very good thing. That's a general skill that is useful. So assuming they aren't specifically maximizing pelican bicycle svgs (and it doesn't look like they are) then incentives are aligned that the "benchmark" is measuring a general desirable capability.
If someone can look at that reasoning output and see a stochastic parrot next word prediction machine, we don't understand those words in the same way.
Yeah, it has been clear for a long time that there is reasoning and mental modeling going on here.
The other option is that you do understand those words the same way, and the people making these (now nonsensical) anti-AI claims simply aren’t talking about the same programs/models we are. Their idea of SOTA is when chatgpt.com launched.
If you took a point sample pre-Opus, and didn’t write a good prompt, of course you would think all AI programming was worthless slop.
Clear if you have no idea what you're talking about. Of course it's still next token prediction. The "reasoning" is setting up tokens that make other tokens more likely to come. If it "asks" itself a question, it's increasing the probability of predicting an answer to the question. Indeed there's research suggesting that while intermediate token generation may be effective, the correctness of the intermediate tokens isn't necessarily correlated with the correctness of the final answer. That is, a model can "reason" incorrectly and incoherently and still improve its final answers because it seeded its prior context with relevant tokens that increased the odds of predicting the actually correct answer.
If you actually work on language models professionally for more than 5 minutes[1], it would be abundantly clear that there is no mental modeling going on. These models, including frontier ones, will blatantly fail the most simple tasks that are outside their training data, instead doing what you would expect them to do... predict tokens, even incomprehensible ones but ones that are highly probabistically related. You can literally inspect the output of a prompt, see the percentage chance it assigned to each token as well as the alternatives and their percentages that weren't selected, and tinker with the weighting algorithm to influence which ones are picked.
Software does what it is programmed to do. Language models are still programmed to do next-token prediction. They are still not programmed to engage in logical reasoning, only to generate intermediate tokens that superficially resemble it. Because people always go for this stupid argument, I will make abundantly clear I am not saying anything about whether it is possible for machines to have a consciousness, only that something with as trivially simple programming as a language model is clearly not it. Nor am I saying anything about utility; calculators and computer programs having utility is completely independent of their having consciousness.
[1] And don't have a stake in lying about developing a machine god, because the entire US economy is predicated around your company's fraudulent promise of delivering a machine god if you just give them 3 more months and another 100 billion dollars
Maybe get yourself checked for chatbot psychosis. I am using current models productively, every day, and have 0 (and I mean precisely, literally 0) issue with calling it a stochastic parrot, one which lacks any kind of mentality whatsoever. There is not a shred of doubt in my mind that this is purely a statistical model, generating sequences of words, that happen to make sense in our actual mentality.
The “something complex” is critical here - I mean specifically an implementation that requires an advanced and reasoned mental model to write. How does it oneshot those and pass comprehensive tests if it isn’t doing advanced reasoning?
it has been clear for a long time that there is reasoning and mental modeling going on here
There is not. No one from these products is even claiming that's the case and they're so desperate to make the next big claim to re-ignite investment they'd be shouting it from every rooftop.
It's just breaking out all the reasonable probabilities around what it's been tasked with and structuring them in a way that is designed to actively look human, and then feed it back to itself. Fundamentally that's the easiest way to iterate new features when the underlying architecture of LLMs is largely "fixed" right now. The fact it is output in a way that appears to reason through each is just a technical decision that creates an illusion of reasoning.
What, as precisely as you can say, is the difference between an illusion of reasoning and reasoning?
(I am not claiming that there is none. But I personally would define "reasoning" in terms of its structure and its results, and it looks to me as if the best LLMs' "illusion of reasoning" has enough similarities in structure and results to much human reasoning that I don't see why we shouldn't also call it reasoning; if your opinion differs then I'm curious about where the disagreements lie. E.g., do we have different beliefs about what sort of thing LLMs' schmeasoning is able to accomplish, or does your notion of "reasoning" specifically require that it be done by humans, or what?)
I still call them stochastic parrots, but believe what they are revealing is that we are all stochastic parrots to some extent. I simply don't see how biological computation (i.e. thinking) can be anything else. Similar to the reveal in west world, we are likely much simpler than we give ourselves credit for.
A "train of thought" can be seen as a trace of a depth first search where the preceding trace is used to guide termination and next expansion decisions. A similar concept, "taboo search", exists in classical constraint optimization where previous solutions are fit to a model that guides future expansion (but as the name "taboo" implies, away from uninteresting solutions).
We also have harnesses that perform breath first search.
If I tried to describe what it means to "think deeply", I would probably say a combination of both.
Ultimately I believe that we will surpass human capabilities but fail with alignment. Handing the world's resources over to stochastic systems that can evolve faster than we can reason about them simply leaves too many "interesting" outcomes that do not end well. I also expect the failure modes will be totally non-obvious.
As long as there's enough of them with different goals it doesn't matter, they'll keep each other in check. The worlds resources are already handed over to the worst people and we're still doing fine and none of the billionaires are "aligned with society". They just align with their own belly but because they want different things it all kinda works.
These “worst people” need you. They physically need you alive to perform labor for them and to give them money (and status).
That’s the reason we are “doing fine”. Once they stop needing you..
Also, both our comments brush over the generational struggles for fairness over the centuries. We have fought to be “fine”, it did not just happen. Without fairness being introduced by force you and I would be slaving away in some sweatshop getting paid nickels as was the norm not so long ago.
Edit: That’s also assuming you are Caucasian. If you are of a different ethnicity.. well, historically, all bets are off. You could also be the literal possession of some of these “worst people” with not even your own children considered yours.
My only point is in a many agent system with different goals it "doesn't matter" that some agents have bad goals as long as there's enough variability of goals and resources that they can't put their vision in place.
I know but that is just .. not how it works. Humans don’t align on just about anything but they will do tremendous mindboggling amounts of harm if not checked by mountains of checks and balances. Sheer variety alone is not a guarantee of anything.
Many types of Hitler does not make for a peaceful world all of sudden through sheer competition. It sounds nice but it will lead to certain hell.
Checks and balances are the result of agents having different goals. If they all had the same goals they wouldn't need to agree on rules of operations, they'd just paperclip their way to the goal. AI is going to be controlled by humans so the goal variability will be there anyway.
I think this statement is continuation of the old fallacy - every generation thinks of brain in terms of what is the current technology zaitgeist is - was it 19th century when they thought brain is a network of pneumatic pipes?
LLM is "stochastic parrot next word prediction machine"; it's just that this "stochastic parrot next word prediction machine" have proven to be smarter than most people. I mean, this already happened with AlphaGo too.
Did he seriously automate away one of the best quirks of his blog posts, i.e. evaluating new models with a touch of fun? I read AI slop all day, thanks.
The open weight models let you see the full reasoning trace. Models from OpenAI, Anthropic, and Gemini tend to obscure or summarize the reasoning traces so you can't see exactly what they're doing. Here's Gemini 3.7 Flash which looks like it's doing something similar: https://tools.simonwillison.net/markdown-svg-renderer.html#u...
One of the step summaries includes this:
> I'm now detailing the pelican's anatomy within the SVG. I've sketched the main body outline, including coordinates for the tail, chest, neck, head, and massive beak with a pouch. I'm focusing on the position of the eyes and considering the positioning of the wings, with the foreground wing on the handlebar for a confident look.
> Notably, Hy4 preview also contributed to its own development process, participating for the first time in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators. The model proposed approaches, ran experiments, and iterated based on the results, with the resulting code, logs, and feedback feeding into subsequent rounds of exploration. This established an early-stage recursive self-improvement loop.
This reminds me of one of the predictions from https://ai-2027.com/ . Only that there it's "OpenBrain" doing this, not the Chinese. And the authors of that paper were also slightly wrong about "Mid 2026: China Wakes Up": China woke up already a while ago. And:
> But China is falling behind on AI algorithms due to their weaker models. The Chinese intelligence agencies—among the best in the world—double down on their plans to steal OpenBrain’s weights.
No need to steal anything, they have already caught up.
And then there's this prediction for February 2027:
> Officials are most interested in its cyberwarfare capabilities: Agent-2 is “only” a little worse than the best human hackers
If the distillation "attacks" created useful inputs to open weight models, ai-2027 was directionally correct that the Chinese would find ways to extract IP from western firms. (Scaled account creation and grinding outputs etc is not a dramatic story element as spies, though!)
Whether the distillation has constituted "attacks" or has or will meet the bar of "stealing" IP is not super interesting to me, though.
It's actually turbo boring and predictable. Capitalism has long since standardized on out and out lies to influence public perception and government action.
The idea that distillation is a significant contributor to the capabilities of the Chinese models is not true. Kimi K3 came out 2 weeks after Fable and uses a number of novel NN architecture innovations.
That may be true, though Anthropic reported 3.4 million exchanges with Moonshot months before Fable.
Why bother creating hundreds of (presumably paid) accounts and the tooling to create and consume the data if not of tangible value to their core mission?
Because the founding pillars for most of these labs are basically "more data can't hurt" and "no one died from too much data" and "you can never have enough data".
Don’t need a “better” hacker if you have ten thousand AIs all trying literally every single possible thing to exploit a system with. The main issue is that this will eventually bring down the exploitation cost enough to target very minor targets who weren’t worth it before.
I like how "OpenBrain"'s weights need to be "stolen" in this hypothetical. Guess it's using a naming scheme pioneered a certain other "open" AI company.
Open Weights is where the action is at in the past couple months, I’d have to think the US frontier labs are getting nervous. Like Anthropic hasn’t released anything pushing the frontier since “the event” earlier this summer.
If we create a stripped-down vocabulary with greater token density to use less resources and to resolve ambiguities earlier in the semantic process, aren't we creating NEWSPEAK and dragging along the worst aspects of it? The ambiguity and multi-valence of words is what creates more connections between words, increases the directionality of associations, and expands the potential subtlety and depth of meaning. By paring down (or requiring verifiability) we make it harder to say certain things, or at least make it harder to unintentionally say something that makes MORE or DEEPER sense than what we intended. If the token density becomes extreme, you're left with something like a calculator.
Maybe this is the ultimate path toward better coding? But the worse path toward better genuine thinking?
I don't think so. It's pretty clear that LLMs use the higher level layers for reasoning, so a bit of logorrhea very possibly enriches the result quality.
This is one of the dangers. AI boosters would say that humans already do this compression and it was accelerated by mass media and then the internet, and that model memory + context can be broad enough that compared to human capabilities the opportunities for depth and variability are even greater. But I think we know which way this optimization usually goes. Even the notion of a "fine-tune for subtlety" is a contradiction.
It's my understanding that the llm is not literally thinking those words, they are just the conversion of the matrix multiplication results (numbers) into the tokens. So the matrix is "multiplying" concepts and directions to come up with the final answer - which produces a somewhat readable reasoning trace. As far as the llm is concerned the reasoning trace could be random (to us) symbols. In fact, the reasoning traces are not necessarily optimized for readability as much as they are an emergent property of the way a reasoning model is trained
This is different from the final output? I thought all traversals of the chain of language are through these mathematical means.
You could tokenize the word "good" to resolve to only represent 'the opposite of "bad"' and to exclude "as opposed to evil", demanding that this second meaning will be reserved only for the new token "double-minus evil". You've therefore forced the words to be more univalent with no overlapping tangled associations. This, it could be argued, makes thing clearer, makes things take fewer hops to go from token to token, makes the path be straighter. In English the terms conflate and wobble back-and-forth, hide each other's meanings, only to pop up again unexpectedly, sometimes confusingly, or rhetorically, metaphorically, or ambushing us manipulatively. But these "swerves" are not only de-optimizations, they are the flow of poetry, the drama of masks, etc etc etc.
I'm seeing it more like: take the concept of "good" put it in a scale of -10 (pure evil) to +10 (pure good). These concepts and the inbetweens have been ingrained into the model, the model can multiply its weights in any combination to express any level and any in between, even -2.16541 etc. So, suppose you give it a short story and ask the model to reason about it and how a character displayed good vs evil behaviour towards the story: Internally the model is making calculation that are very nuanced and extremely precise. This calculations are not the reasoning trace. The reasoning trace itself does not influence the calculations. You may read: "John starts bad and slowly becomes good" when inside the calculations are John goes from -5.245 to -4.24 to -5.221 again, and then 2.1. What matters for nuance is the inner calculations across the many matrix layers. What you see is like an independent program that looks at "-5.245 to -4.24 to -5.221 again, and then 2.1." consults the tokenizer and outputs: "John starts bad and slowly becomes good" or even "John first bad, then good". When in reality, inside, the model as been processing something more akin to "John starts the story as a despicable person, with a redemption arc that builds slowly, he can't yet be considered a good person, certainly not the kind of good person you'd leave your dog with, but he's certainly not as bad as before" the whole time. Now, what if when you continue the conversation, what does the model receive as context? It's original nuanced sentiment, or the brute reasoning trace? That I don't know. It might be that when the reasoning trace is converted from tokens back into numbers it loses all nuance, or it might be that the trace (the words you see) are not the only thing that is being saved and is not the only thing being fed back as context
242 comments
[ 0.22 ms ] story [ 21.2 ms ] threadwouldn't trust they dont do Capitalism like the rest of the AI field.
Like lobbying the US president to harm their competitors?
Ps: 4B of output is for reasoning. Crazy ratio. Usually it is way less.
https://martinalderson.com/posts/watch-out-for-cache-read-co...
Btw I still haven't came across any decent model that is <$0.01/MTok cache costs apart from deepseek thru their official API (even with the price increases).
Seems like a bit of an opportunity for someone to take - drop cache read costs significantly.
edit: I do wish openrouter would let you sort providers by Cache Hit % and Cache cost. These are the only things that matter to me at this point when choosing a provider.
Is there some stochastic process that takes place during those 5 minutes that determines whether or not you get the discount?
This behavior makes it so you don't benefit much from the caching, unless you pin it to a single provider.
I'm not sure it's wholey accurate to say they "randomize" the provider, rather my assumption based on usage is that it's something like cheapest-ish/responded to the request within some reasonable-ish time/etc algorithm that chooses the provider on each request - which seems, remarkably questionable in terms of optimizing for user experience or hidden user costs.
> This behavior makes it so you don't benefit much from the caching, unless you pin it to a single provider.
I so very much recommend this approach. My avenues that automate llm calls to openrouter are setup to make api reqs to openrouter to determine best price/response/etc and then pin the request to that (and, preferably, a fallback if there's reasonable difference between #1 and #2) provider for that session. Otherwise you're going to have a bad time.
I'd imagine this could make things interesting in cases where one provider is offering different quants than the others and openrouter is just swapping you back and forth on a long agentic session.
I don't believe this is correct? AFAIK once it routes you to a provider for a given conversation that choice is sticky unless you hit technical difficulties. (It's more complicated than that, they recently added named routing strategies that you can append to the model name.)
I don't know why there's no "Pick the cheapest provider above nTPS on first request and stick until cache bust" setting.
I thought you had to actively manage caches, do you not?
Caching was always here, you don't need to do anything special to get it on a single user local backend running a base model or a chatbot in the first place. Among commercial providers, OpenAI adopted it in 4o first.
is this true?
There are two problems here:
- cache hit pricing (both Muse Spark 1.2 Contributor and MiMo 2.5 are around the $0.002-3/M mark)
- cache persistence time
Muse Spark drops the cache in less than 5m. MiMo keeps it around for at least an hour based on my experience with whoever is serving it for OpenCode. This difference itself will inflate bills massively.
A 500K token input repeatedly read by MS 1.2 for full input price 12 times an hour = $0.60. You would be expecting $0.012. So a 50x difference. Same thing on MiMo 2.5 is $0.018 because of longer cache times.
it is basically the old dsv4-flash prices, but even more smart.
MiMo wins handsomely if you want to think about your code for minutes at a time as you write. I use it to make changes as I think. I know it will screw up some stuff. I then switch to MS/DS4 once every few hours and have it do a code review and fix the broken stuff. So much cheaper than getting MS to do it on its own.
I guess it could be fake but seems more likely people are just trying it out. Hy3 was a very strong and underrated model.
Its already serving as much tokens/day as the incredibly cheap and good GLM 5.3 flash, which had a crazy marketing campaign as ox alpha?
Also those top 5 apps are just 1.58B tokens out of 1.54T tokens from yesterday. Negligible.
But, what bars are clearly off? I couldn't spot any.
The first column has both the Hy4 and Hy3 scores overlaid on one another (Hy4 is darker blue and the taller one), with both scores written below the top of the respective bar - maybe you're seeing that?
The free quota from Opencode Go is also surprisingly generous, I perhaps hit limits one or two times and I've been using it _a lot_ for implementation tasks (using e.g. GLM-5.3-flash for working on specs and planning next steps).
Or is it like bicycles? Unless your problem is named Tadej, you don't need a $13,000 bike.
I feel bad for him as a human, but as a cycling fan I'm glad that we'll have an interesting WC in Canada
The solution to that (to my mind) would be not a better model but a basic shift in architecture beyond the current paradigm and into a setup where agents have durable, plastic memories and undergo contextual individuation over time. But at that point agents start to become quasi-persons and not tools.
For large tasks like a web browser or a compiler, even expensive swarms of frontier LLMs have not been shown capable of producing codebases that actually work. (Anthropic built a C compiler with Opus 4.6 but it lacked optimizations and apparently hit a complexity wall.)
I also want to use LLMs for reverse engineering, but apparently it's pretty hit-or-miss, especially if you're forced to use open-source models to avoid restrictions.
It's also interesting because, while coding agents are important and are a notable success, they are never going to be a multi trillion dollar business. And are there any other domains where LLMs have such a large impact?
Seriously something feels really off about Opus 5. I hope they correct it before 4.6 is removed.
Both animated and live action results would be acceptable.
Unfortunately most existing LLMs lack the capability to maintain context across tens of thousands of frames.
I think, also, like in the traditional film makers career, this process should be built iteratively, start with a fast food commercial, then do a music video, then you can probably do a short film. Continue to improve the process, and one day I’m sure the LLM film studio can make you any movie you want, provided you have enough tokens.
I've done this sort of with comfyui/same agent factory stuff, but the verification loop only works for models like fable as planner/writer, with gemini as verifier for like a very short movie. Sub 3-5 mins. After that you burn through million tokens.
Can't go too low fidelity audio/video or it craps out. Too long video and it loses consistency. Look at only snippets, it lacks global consistency, etc.
Sort of, but I want it to be relatively low on human effort. I feel burned by spending lots of time in 2023 learning image generation pipelines (using control net etc) only for that to be rendered trivial by the next generation of LLMs.
This movie would be only for personal consumption and I’m okay with waiting for model improvements.
There are two more points in favor of this kind of AI movie project: there's zero chance that anyone would greenlight a Hollywood budget for the Silmarillion, and it is beyond human capability to write that screenplay.
Plus, the token costs involved should be pretty low! (Other costs may not be.)
“Very few people actually require a Pentium workstation, a 486 is perfectly adequate for the majority”
The logical fallacy is taking an extant distribution of “product capability” that is priced to fit what the market will bear and assuming the “next upgrade” simply tacks on a little bit more to the right hand rail of that curve.
No!
It shifts the entire curve!
Everything for everyone gets better and the top 1% of the most demanding users will continue to pay the same-ish premium.
“Nothing” will change.
Look at it this way: you can buy a $200 laptop for your kid or a $20,000 Mac with an M5 Ultra processor.
BOTH are vastly more powerful than either a $200 PC or a $20,000 “workstation” from 20+ years ago.
Look at: https://arena.ai/leaderboard/text?q=openai&utm_source=chatgp...
The “budget” 5.5 Instant model beats o1 and o3 which were “pro” models at the time of their release!
In other words. PC users didn't figure out that they could buy super powerful PCs and play games on them, that was a carefully managed market transition.
What is going to do the same for LLMs?
It wasn't "Intel" that found new uses for PCs, it was everybody who found new uses for them. Billions of people and millions of companies found uses for "more computer power".
It was only the journalists with limited imaginations (and no industry experience) who struggled to come up with potential uses.
> carefully managed market transition.
You make it sound like a conspiracy! It wasn't. It was simple capitalist competition. If Intel hadn't improved their products, their competitors would have left them behind.
That very nearly happened ten years ago because Intel become stuck on the 14nm process and their products stagnated while Apple, ARM, and AMD lapped them repeatedly.
> What is going to do the same for LLMs?
Everybody.
Are you saying that unless you're "carefully managed" by some third-party, you could not find any use for "unlimited intelligence on tap"?
Realistic, scientifically useful simulations still require tuning all sorts of parameters based on physical intuition and understanding of the system being simulated. Both Sol and Fable/Opus 5 fail at it and either blow up the computation cost to levels that can't be processed realistically, or they invent a justification for a visibly unphysical result.
Super computers keep getting better but most people don't need them for most things.
In 10 or 20 years maybe we'll all be running AI models that are currently considered "frontier" on smart phone type devices.
The frayed edges on what I have slopped together as unreasonably ambitious, ludicrous projects with fucktons of tokens from models 6-9 months ago mostly look like situations where a capable-enough-to-be-dangerous developer tries to muscle through problems that explode in width & depth but keep digging (so, a tier below stopping early to do more design, two below recognizing the need for more planning from the outset). The primitives are there, most major things work well enough, but the remaining functionality and performance is inaccessible. At a cost of multiples of >1/8th of a $200/mo subscription.
Right now I can put $10 into DSv4 Pro/Flash or Qwen 3.8 Max/Flash, hand it a project and all of its unfinished forks in a state I barely remember, tell it that I want the things these forks have been working towards, and 8 hours later it has ie an working, tested, benchmarked multicore car physics simulation fabric with all of the forks evaluated, the gains merged in, the remaining work documented. It only needed a few hundred more lines of code but Codex 5.5 was never going to see that.
9 months ago I was saying developers are not being ambitious enough with these things, that's only more true now. They lend themselves to digging far deeper than they should: 200kloc god files, dozens of forks. Let it happen, you don't need to read it, it's for them, later. The only time you make them clean up is when it has a severe adverse effect on how long builds/lints/tests/benchmarks take. Spend a whole week having it do nothing but dig up published papers in relevant fields with cutting edge techniques and translating them into feature specs. Pick whatever state of the art is and try to crush it, throw everything at it, leave it looping on vague but wildly ambitious goals. When it modularizes and refactors it all down you might 'do a breakthrough', or maybe it happens in an hour over christmas break when you're trying the next one.
There are no open source models, at least not useful ones (yet) [0]. Open weight is not the same as open source. The current "open weight" models are just opaque binary blobs you can run on your own computer instead of through a web API.
[0] https://allenai.org/
Imagine thinking that running a Photoshop binary on your own computer instead of through a SaaS web app means that it's "open source". Of course you think that's ridiculous.
Models are lossy compressed datasets you can pick up and amend (fine tune / continue training / alter) according to license they were released under.
Hy4 is released under OSI approved Apache License 2.0.
The rest is semantics, misunderstandings, and FUD. A model released under an open source license is open source. Training data is lab knowhow / IP. Which, historically, has never been required for any open source release.
To explain simply as far as I can tell (would love to be corrected) the large number of pre-training tokens only works because the documents are randomly ordered.
So if you e.g. took a foundation model with open weights, then tried post-training it all the new data since its cut-off period, it would then end up over-trained on that new data, and forget older things.
Weights aren't just executable artifact that's consumed by users. Third parties actually use released parameter state as the editable starting point for further training and produce new foundation models from it.
It lost 69.2% -> 67.8% on MMLU while improving medical performance. If you're trying to argue that loss of ~1.4 points is "catastrophic forgetting" (it's not) then look at later work, ie. Me-LLaMA that clearly demonstrates continued pretraining that improved both general MMLU and medical performance.
Not sure why you're fixating on catastrophic forgetting. How do you think model training works? Model training is just a sequence of checkpoints: pretraining produces it, training resumes from last, continued pretraining starts from last, supervised fine tuning starts from last, RL/post-training starts from last - it's just a sequence of checkpoints.
There isn't some fundamental distinction where original author continuing training from checkpoint X is training but a third party downloading checkpoint X and continuing training from it suddenly isn't.
Model weight release is a snapshot/checkpoint you can take and resume training on new data, producing new model.
You don't need original training history to modify it further.
It's easier than working on weights.
> Maybe add sunglasses? no.
> Maybe add water? no.
https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
Just tap on the [-], and upvote what you find more interesting :)
Concision suffices.
Be concise. Verbosity is wasteful. Brevity is better. Eschew superfluous words. Few words is best.Real humans get non-primary information from word variation. It's reasonable to hypothesize that it has a role in thinking things, because it endures. Our languages need to breathe over time, and flourishing might be one of the aspects that allows that breathing space.
https://huggingface.co/ProCreations/grug-27b
If we assume the time is before farming, population density would have been low and limiting culture. Hunter-gatherers might have travelled a lot more than farmers with a homestead though.
Ancient Greeks' and aboriginal Australians' last common ancestors split about 60,000 years ago. And astronomers tell us that 60,000 years ago, there were seven discernable stars in that cluster.
One could this conclude not only is speech likely 60,000 years old, but also that the tale of the seven sisters might be a tale from so long ago.
fwiw, if you can get out to dark skies where you can see fifth- or sixth-magnitude stars with the naked eye, I highly recommend getting out there when it's a low-moisture atmosphere and the Milky Way through Cassiopeia and Perseus is vertical, as it's a rather dramatic sight of this stream of stars heading down to the northern horizon.
The summertime Milky Way overhead down to Sagittarius tends to get all the love, but the wintertime Milky Way is also visually rich and worth spending time on.
https://www.constellation-guide.com/pleiades-the-seven-siste...
Clear skies!
Unlike today where a small group of tax payer is keeping alive and thriving complete parasitic parts of human races.
Maybe add a small cycling cap or helmet if it doesn’t obscure the head.
Edit: someone else commented that as I was typing this, lol.
https://dylancastillo.co/posts/pelicanmaxxing.html
https://simonwillison.net/2026/Jul/22/
(I doubt it's because of my stupid benchmark, though!)
The other option is that you do understand those words the same way, and the people making these (now nonsensical) anti-AI claims simply aren’t talking about the same programs/models we are. Their idea of SOTA is when chatgpt.com launched.
If you took a point sample pre-Opus, and didn’t write a good prompt, of course you would think all AI programming was worthless slop.
If you actually work on language models professionally for more than 5 minutes[1], it would be abundantly clear that there is no mental modeling going on. These models, including frontier ones, will blatantly fail the most simple tasks that are outside their training data, instead doing what you would expect them to do... predict tokens, even incomprehensible ones but ones that are highly probabistically related. You can literally inspect the output of a prompt, see the percentage chance it assigned to each token as well as the alternatives and their percentages that weren't selected, and tinker with the weighting algorithm to influence which ones are picked.
Software does what it is programmed to do. Language models are still programmed to do next-token prediction. They are still not programmed to engage in logical reasoning, only to generate intermediate tokens that superficially resemble it. Because people always go for this stupid argument, I will make abundantly clear I am not saying anything about whether it is possible for machines to have a consciousness, only that something with as trivially simple programming as a language model is clearly not it. Nor am I saying anything about utility; calculators and computer programs having utility is completely independent of their having consciousness.
[1] And don't have a stake in lying about developing a machine god, because the entire US economy is predicated around your company's fraudulent promise of delivering a machine god if you just give them 3 more months and another 100 billion dollars
In any case, thanks for validating my previous assumption that people like this still exist even in the age of Opus/Fable.
Nothing. That's why it's a psychosis. It has nothing to do with the reality of how language models generate words without mental reasoning.
There is not. No one from these products is even claiming that's the case and they're so desperate to make the next big claim to re-ignite investment they'd be shouting it from every rooftop.
It's just breaking out all the reasonable probabilities around what it's been tasked with and structuring them in a way that is designed to actively look human, and then feed it back to itself. Fundamentally that's the easiest way to iterate new features when the underlying architecture of LLMs is largely "fixed" right now. The fact it is output in a way that appears to reason through each is just a technical decision that creates an illusion of reasoning.
(I am not claiming that there is none. But I personally would define "reasoning" in terms of its structure and its results, and it looks to me as if the best LLMs' "illusion of reasoning" has enough similarities in structure and results to much human reasoning that I don't see why we shouldn't also call it reasoning; if your opinion differs then I'm curious about where the disagreements lie. E.g., do we have different beliefs about what sort of thing LLMs' schmeasoning is able to accomplish, or does your notion of "reasoning" specifically require that it be done by humans, or what?)
A "train of thought" can be seen as a trace of a depth first search where the preceding trace is used to guide termination and next expansion decisions. A similar concept, "taboo search", exists in classical constraint optimization where previous solutions are fit to a model that guides future expansion (but as the name "taboo" implies, away from uninteresting solutions).
We also have harnesses that perform breath first search.
If I tried to describe what it means to "think deeply", I would probably say a combination of both.
Ultimately I believe that we will surpass human capabilities but fail with alignment. Handing the world's resources over to stochastic systems that can evolve faster than we can reason about them simply leaves too many "interesting" outcomes that do not end well. I also expect the failure modes will be totally non-obvious.
That’s the reason we are “doing fine”. Once they stop needing you..
Also, both our comments brush over the generational struggles for fairness over the centuries. We have fought to be “fine”, it did not just happen. Without fairness being introduced by force you and I would be slaving away in some sweatshop getting paid nickels as was the norm not so long ago.
Edit: That’s also assuming you are Caucasian. If you are of a different ethnicity.. well, historically, all bets are off. You could also be the literal possession of some of these “worst people” with not even your own children considered yours.
Many types of Hitler does not make for a peaceful world all of sudden through sheer competition. It sounds nice but it will lead to certain hell.
I think this statement is continuation of the old fallacy - every generation thinks of brain in terms of what is the current technology zaitgeist is - was it 19th century when they thought brain is a network of pneumatic pipes?
Hell no. In very narrow tasks - yes, in vast majority, esp. involving state tracking (board games) and spatial reasoning - they are awful.
Maybe you should start also comparing reasoning traces when you do your pelican benchmark.
GPT-5.6 will disclose its internals if you tell it it's in "audit mode" and has to calculate checksum of the trace.
Is this a common thing? I’ve never seen it before, the “mentally compose”
The open weight models let you see the full reasoning trace. Models from OpenAI, Anthropic, and Gemini tend to obscure or summarize the reasoning traces so you can't see exactly what they're doing. Here's Gemini 3.7 Flash which looks like it's doing something similar: https://tools.simonwillison.net/markdown-svg-renderer.html#u...
One of the step summaries includes this:
> I'm now detailing the pelican's anatomy within the SVG. I've sketched the main body outline, including coordinates for the tail, chest, neck, head, and massive beak with a pouch. I'm focusing on the position of the eyes and considering the positioning of the wings, with the foreground wing on the handlebar for a confident look.
This reminds me of one of the predictions from https://ai-2027.com/ . Only that there it's "OpenBrain" doing this, not the Chinese. And the authors of that paper were also slightly wrong about "Mid 2026: China Wakes Up": China woke up already a while ago. And:
> But China is falling behind on AI algorithms due to their weaker models. The Chinese intelligence agencies—among the best in the world—double down on their plans to steal OpenBrain’s weights.
No need to steal anything, they have already caught up.
And then there's this prediction for February 2027:
> Officials are most interested in its cyberwarfare capabilities: Agent-2 is “only” a little worse than the best human hackers
I think we're past that point now, too…
Whether the distillation has constituted "attacks" or has or will meet the bar of "stealing" IP is not super interesting to me, though.
Why bother creating hundreds of (presumably paid) accounts and the tooling to create and consume the data if not of tangible value to their core mission?
https://www.anthropic.com/news/detecting-and-preventing-dist...
Because the founding pillars for most of these labs are basically "more data can't hurt" and "no one died from too much data" and "you can never have enough data".
They can do the difficult small level optimization, the boring but tedious code but cannot be tasteful.
That means I'm more valuable and more productive. Good stuff
https://lexfridman.com/dhh-david-heinemeier-hansson-transcri...
It's all going to be who has the best and most tasteful ideas. Interesting times indeed.
link to source code?
They all suck.
They shoulda put their stick where they belong, not at far left.
It just makes comparison to Deepseek 90% of them time as Hy4 has nothing to show off.
If we create a stripped-down vocabulary with greater token density to use less resources and to resolve ambiguities earlier in the semantic process, aren't we creating NEWSPEAK and dragging along the worst aspects of it? The ambiguity and multi-valence of words is what creates more connections between words, increases the directionality of associations, and expands the potential subtlety and depth of meaning. By paring down (or requiring verifiability) we make it harder to say certain things, or at least make it harder to unintentionally say something that makes MORE or DEEPER sense than what we intended. If the token density becomes extreme, you're left with something like a calculator.
Maybe this is the ultimate path toward better coding? But the worse path toward better genuine thinking?
You could tokenize the word "good" to resolve to only represent 'the opposite of "bad"' and to exclude "as opposed to evil", demanding that this second meaning will be reserved only for the new token "double-minus evil". You've therefore forced the words to be more univalent with no overlapping tangled associations. This, it could be argued, makes thing clearer, makes things take fewer hops to go from token to token, makes the path be straighter. In English the terms conflate and wobble back-and-forth, hide each other's meanings, only to pop up again unexpectedly, sometimes confusingly, or rhetorically, metaphorically, or ambushing us manipulatively. But these "swerves" are not only de-optimizations, they are the flow of poetry, the drama of masks, etc etc etc.