The astonishing thing with the Halo franchise is that 343 kept being given chances to go again, and again, when it clearly just wasn't working. By the time they get round to releasing a great new Halo people that loved the original will be taking out their pensions.
It cannot be understated the extent to which the console industry, including third party publishers and devs, needs Sony and Microsoft to both stay competitive. In that light the drastic moves being done at MS are long overdue, but maybe not sufficient. It's also an open question if consoles (outside of the Nintendo universe, which is a very different audience) have a future at all.
Halo used to be what everyone wanted (both to play and dethrone), but now it's sitting in the corner eating glue. Moving it under Activision is certainly an extreme measure, but if it means everyone who touched on it in the past 15 years can't anymore, I'm listening.
But humans have been confusing pattern matching against known solutions for intelligence for a hundred years!
Seriously though, it ends up looking like that. To take a stupid example a couple of weeks ago I asked an agent to look at porting my hand written WebGL renderer (+ shaders etc) to WebGPU. It estimated a human would take 6-10 weeks, and I would agree. (Which is why I hadn't done it). 24 hours later it was deployed and live. This is classic tedious, difficult, low level if quasi mechanical work (rather like cracking an enigma message), and LLMs absolutely fly through it.
Time estimations by LLMs are hilariously incorrect all the time. It estimates very simple things that a human could do in an hour to take days or weeks and other things that are genuinely tedious and time-consuming it estimates taking a few hours. LLMs have no understanding of time and no world model that even allows them to make correct time estimates. They will always fail to provide decent time estimates unless the task is well-known, in the training data and they can extrapolate that with a simple math script.
My favourite story is of an elderly coworker with no AI experience who innocuously asked GPT-5-mini (at the time his framework's default model) to translate a 50 page engineering spec, and it kept him at arm's length for about a week about how that task would take just another 24-36 hours more. He kept asking like, you finished yet? and it just made up excuses, "so sorry, I got distracted, give me one more day", and he went "please finish, I need this", and 5-mini came back "I totally understand -- let me get to work immediately, I'll report back ASAP", end of conversation. He got annoyed but never even suspected anything wrong, because this is the kind of conversation he's used to. I wonder how long this would have continued, if I hadn't intervened by chance.
You do understand this is intentionally trained into recent models for marketing purposes? "Wow, it saved me months of work in a day! This is the most amazing technology ever!!!!"... is what it intends to evoke by underpromising and overdelivering. I routinely have it helpfully suggest it will take something like "three engineer-months" to do something I do by hand without any LLM assistance in a day. The estimates may be accurate if you have literally never touched a computer in your life before and are starting to learn from there.
In the games industry I was tech lead of teams of hundreds of devs and had to deal with their estimates of this sort on a daily basis. 6-10 weeks for a total renderer rewrite is on the low end.
As an indie dev who built their own WASM-capable engine that I've shipped in real games, I've built both WebGL and WebGPU renderers from scratch myself in significantly less time than that. Sounds like typical corporate dysfunctionality. If you have hundreds of devs you're going to get bogged down by having a share who spend 90% of their time at the company on Reddit, another share who write actively bad code, and then maybe 10% of the employees who have a clue what they're doing dealing with the overhead of communication, meetings, other people not upholding their assigned responsibility, etc. slowing them down 10x what they could actually do.
0% chance, without LLMs. WASM-capable engines would require years if not decades to build, even with expert level knowledge. See Jonathan Blow who has been working on his engine for a decade now, and he's arguably of the most talented engineers who ever lived.
Jonathan Blow is writing his own language as well, and is also already successful enough that his work is just a hobby he can take at any pace with no urgency. A WASM engine is really, really not that difficult. At its core, you need rendering to a canvas, audio, keyboard/mouse/gamepad handling, asset loading/file saving, and an update loop. Writing this code is mostly not different from writing code in other languages, since you are literally writing other languages that happen to compile to WASM, the only differences being that you need your one-time WASM toolchain setup, some JS glue interop (which is not really different from needing C interop for native engines written in languages other than C), and to be aware of browser pecularities regarding file access, update loop, threads, etc. which you also have to deal with if you ever wrote a JS game anyways. After you have those core elements in place, everything above that is isolated in a game/engine logic abstraction layer that isn't any different from writing native code.
If you want to place a bet on it, we can do a $10,000 bet in escrow contingent on myself implementing a well-specified WASM engine from scratch on stream without LLM usage in a month. I would love an opportunity to demonstrate how wrong you are. That said, rather than taking your money, I could also just share a streamer's content with you[1]. He implemented 3D web rendering with no dependencies in a 20 minute lecture, and it would take 10 minutes if you were seriously focused on doing it quickly. Sure, it was rudimentary pure JS rather than WASM, but really consider whether you think this 10 minute exercise couldn't be done in a language that compiles to WASM with 160 hours, while including the other hardware/OS-layer abstractions aforementioned. On the other hand, please do take me up on my offer. You said it's a 0% chance, after all -- surely you don't want to pass up on the easiest $10,000 of your life...?
Even those of us that are pro AI need to acknowledge this is the current reality.
The smartest humans now need to move to being less concerned about status games among humans and more with how to provide value to a mix of intelligent machines and humans. i.e. if you're starting an SaaS in 2026 you better be assuming half your revenue is going to come from machines acting by themselves.
The whole "multiplatform" portability business has been one of the most conspicuous victims of the agent wave. Porting is now primarily a QA exercise, not a development one. (Development too, generally, is heading that way).
The problem all these products have is these blog posts that describe in enough detail what the patterns are, out of necessity, also provide the information for the product to be cloned into a form that is then more dynamic and malleable for use by other developers.
It's clearly existential for JetBrains, and it's revealing that the opening goes for the fear with "but it hasn’t changed what it costs to be wrong" while ignoring that this hasn't exactly been a major concern for most software people for a scarily long time, including JetBrains themselves.
The Claude models definitely felt more susceptible to moods, like you could leave them for a few hours, come back and it suddenly was unable to do things which it was doing just earlier, which tellingly is never an experience I've had with an open model.
Honestly I lost patience with Anthropic both clearly messing around with things like this and their agitation over regulation. They aren't good actors, and quite why so many blindly trust them with their company crown jewels is a mystery.
I think it should, as you dont need to use the encoder layer on the new tokens, you just read the embedding from the cache. that's why cache reads are cheaper
It's supposed to be for token optimization (https://code.claude.com/docs/en/prompt-caching), but are people experiencing degraded performance when you let Claude Code sit for hours/days and come back?
Sun commercial sales reps, at least in Europe, were astoundingly slimy.
I had a colleague where almost his whole job consisted of beating them in sheer sliminess, and at this he was good. I don't think we ever paid 30% of the prices they'd start at, but the process of achieving this was disgusting, and there was still the sense everyone was being screwed.
Yeah, and it wasn't just the "call for pricing" game, it was having to deal with monthly "touching base" calls which were mostly an opportunity to push corporate hospitality and other not-quite-inducements.
Whereas Dell just gave us a flat 30% discount on list pricing, and were clear in advance about what level of annual spend would trigger an increase to 35%. We'd order online and our sales rep would phone a few hours later to confirm. Other than that, they'd initiate contact maybe once a year when Dell refreshed their server or storage product lines and would offer to either send some bumpf or arrange for some demo kit.
Dealing with them was so much less hassle all round, you never got covered in slime, and you never felt like they were on the verge of offering you bribes.
Uh, as in "excessive paperwork", like brochures/leaflets [1] about the new products? Not a native speaker, had a parse error I'm just trying to resolve ...
Yep, a mix of glossy brochures and things like white papers from Intel about their latest Xeon 3 architecture or whatever. They tended to lie about on someone's desk for 6 months before being chucked out...
(and I'd had no idea about the "bum fodder" derivation, so thanks!)
Yeah that fits. I was in London (though not the City) and it was definitely that sort of scene.
Thinking about it, we moved to Dell as part of the run up to IPO, and I suspect had we not done so Sun would have been much harder to deal with afterwards.
What I recall from that time is Sun were very happy they'd sold a huge pile of servers to be the core of Interlink, and a whole load of people went off to work on that.
It's interesting how the negative examples all look like mobile game ads.
Which isn't an accident - mobile game ads are tuned studiously on consumers actual preferences and not the preferences of designers. i.e. the people that are likely to spend actual money on the product/service being advertised to prefer the ad being portrayed here as bad.
Bluntly, being able to sidestep the whims and fashions of graphic designers is proving to be a major feature of the AI revolution.
Surely this implies these LLM generated proofs require the LLMs to have mathematical intuition . . . and honestly I don't think many people believe that, and rightly so, certainly not in the way Poincaré was on about.
Maybe it's been done, but I'd like to see an LLM recreate Euclid from questioning without having seen it during training.
> Surely this implies these LLM generated proofs require the LLMs to have mathematical intuition . . . and honestly I don't think many people believe that
Yes, I believe that, it's part of what I was implying (I believe the LLM weights have some internal representation of math in the same way brains do that allow them to produce proofs)
I think we differ on what "mathematical intuition" is then. I've seen people that do well in undergrad math degrees simply by massively memorising things and learning how to join them up to some level of degrees-of-separation, but seemingly completely fail to understand, for ezample, why even calculus is how it is. Because they are able to regurgitate the results and "produce proofs" this is never questioned.
The Euclid example also shows my bias towards spatial intuition of mathematical concepts (which is deeply unfashionable) but also exposes exactly where at least current LLMs break down; they do the symbol based pattern matching version, but they cannot leap outside of that, at least today.
This feels very related to the issues re: the presence or absence of world models in LLMs. Insofar as they have world models (or "intuitions"), these would seem to have to be primarily verbal-linguistic (or symbolic, when using math). LLM world models are not likely (currently) very spatial, in contrast to e.g. V-JEPA-2 models, which likely do have some basic spatial models (and perhaps "intuitions").
> I've seen people that do well in undergrad math degrees simply by massively memorising things and learning how to join them up to some level of degrees-of-separation, but seemingly completely fail to understand, for ezample, why even calculus is how it is. Because they are able to regurgitate the results and "produce proofs" this is never questioned.
If you want to catch them, surely you can find proofs they aren't able to produce.
I used to be a game dev, and one of the interview questions someone came up with consisted of working out the surface area of a variant of Menger sponge to some given level of depth. The bifurcation for people that could do this vs those that couldn't was incredible, and did not follow obvious trends for academic achievement. (The same interview also included the gem "How wide is a pointer?" which also catches a frightening number of people).
I don’t understand this. You seem to look down on people who put in effort and instead celebrate those who have natural talent. Like the people the people are somehow cheating by studying. Seems backwards.
Intuition also takes effort, just of a different type and quality. It's the difference from mindlessly applying memorization and simple inference/recombinations (basically delegating to our own internal LLM) versus deeply contemplating the meaning of certain mathematical objects, looking for intuitive analogues etc. I would say the second takes more effort and seriousness. The first is more like brute-forcing a problem.
I think it's actually the opposite. Intuition is the only thing LLMs can do, which is why they are prone to hallucinating when they can't validate their intuition against reality.
There is an idea that human intuition, expertise, and critical thinking are largely pattern recognition. When you encounter a situation, your brain gives you a plausible starting point, based on what it has experienced before. You then continue with explicit reasoning, which is slow and inefficient, and try to validate your ideas. The more relevant the patterns you have learned are to the situation, the more likely you reach a useful conclusion.
LLMs are largely the same, except that they cannot learn from experience in normal usage. And except that they experience the world only through symbolic data, while the human brain has access to plenty of sensory data.
Kind of. First of all, reasoning LLMs can also do (a form of) reasoning. Second, yes you could say LLMs form a sort of intuition, but its domain is the space of human-produced text. It only translates to real world intuition to the degree that those intuitions make their way into the corpus the model has been trained on. You could say that, when it comes to anything other than textual prediction, their intuition is secondary, a reflection of a reflection, so it will always lag behind that of humans.
My hunch (or intuition, hah!) is that intuition is an instinctive mental shortcut required to navigate large problem spaces that can’t entirely fit into our heads.
Maybe LLMs do not need intuition because they can scale their “cognitive capacity” with hardware and brute force their way through these problem spaces.
> Maybe LLMs do not need intuition because they can scale their “cognitive capacity” with hardware and brute force their way through these problem spaces.
My view is that is certainly true of smaller LLMs but becomes less true as they scale up.
To quote the parent bananaflag in a sub-comment:
> I believe the LLM weights have some internal representation of math in the same way brains do that allow them to produce proofs
I think as the sort of spare space adjacent to pure language processing in LLMs grows the probability of the sort of reasoning bananaflag is getting at (or spatial reasoning, or anything else) emerging in that space grows enormously.
One of the questions for AI development over the coming months or years is going to be if deliberately cultivating the architecture of those sub models for specific reasoning types beats any emergent reasoning mechanisms or not.
Hmm we may be talking of two different interpretations of intuition here. I agree that LLM weights contain representations of abstract concepts, as a lot of prior research has shown. This surely includes Mathematical concepts.
But to me that is analogous to what human brains do, and a bit different from intuition. I think of intuition as “heuristics”, typically developed through experience, that may link seemingly unrelated concepts via vague, hard-to-define associations, but which let us make mental leaps (or shortcuts) while reasoning. (Maybe analogous to System 1 / 2 thinking.)
On the other hand, LLMs can do both: build “intuition” from patterns in data AND brute force a huge amount of potentially unrelated concepts. This gets fuzzier when we realize that even these “concepts” themselves are gleaned from patterns in data! But my point is we necessarily have to take shortcuts to scale, whereas machines can scale with hardware.
This is of course a layman theory! But it could explain why these models are progressing so fast.
Yes, in my case "intuition" comes a lot from visualizing things spatially, manipulating them, and being able to capture their properties in equations/proofs, and it's that which is (currently) conspicuously missing when dealing with LLMs. (And may yet appear with world models).
With the alternate view of intuition that many of you are describing it is clear LLMs are somewhat either there or heading there now.
This is an intriguing observation! LLMs were famously bad at spatial reasoning, until Astra which apparently has a huge improvement. I wonder if that has any bearing on the recent jump in Mathematical performance?
One thing that struck me from Dario's last podcast with Dwarkesh was that he said training LLMs on a diverse set of tasks does not make them better just at those tasks, but they get better at unrelated and other tasks overall. What you described could be a concrete example of how that dynamic works!
Your unstated major premise here is that their intent was to make a universal statement about how proofs work and not just talking to humans about how they teach humans.
Why? The entire subject of conversation is triggered by things which are not humans producing proofs.
If it's possible for a machine to produce a proof without intuition then clearly a human could also do it too. (And in fact I'd argue I've seen many people like that, simply very good at pattern matching over memorised items).
Because regardless of the point TFA is making, that interpretation makes less sense for the specific comment. It doesn’t fit with the immediate context, which was a response to a thoughtful comment about how humans do math. And it requires assuming a math professor doesn’t understand a very basic and obvious thing about their area of expertise.
That doesn’t really read as good faith engagement in the discussion. At best, it reads as being so AI pilled that you can’t even fathom that others might want to have a little side discussion about something other than AI.
That math prof was talking about his stance in discussion between mathematicians long before AI.
Plus, I studied math, I am from that environment. His description matches how math is done by people.
People who are good at pattern matching and memorize are, frankly, shit mathematicians. They are find in fun culture around math, but rarely in actual math. They cant really do it as science.
reply