> people only look at the form and ignore the substance of the content.
This is largely because most people are incapable of thinking critically and evaluating substance. So they regress to indirect markers of good writing, and the whole thing becomes a game of signaling.
The rot is so deep, many popular resources on writing focus more on style and presentation than focusing on the substance of writing. And even when best-intentioned, all this does is make poor writing a bit more palatable by giving it a veneer of polish.
The core thesis of this essay is reminiscent of the Lisp Curse [1] / Bipolar Lisp Programmer [2].
It's been a few years since I read these, but if I recall the argument there, it was that Lisp makes it so easy to build stuff and scratch exactly your own itch, that there's no real strong push for lisp programmers to come together and collaborate to build non-trivial and general purpose artifacts. And that is why the landscape of public lisp software is poorer as a result, compared to languages which demand much more effort to get anything substantial done.
Armin seems to be making a very similar point about AI coding.
I know nothing about you or your code, but till you make your code + workflows public and have all of it reviewed / critiqued (so that other can replicate the performance), it's hard to accept it as a serious case study.
For all we know, this could just as well be a severe case of Dunning-Kruger.
Imagine I want to attend a conference in a different country. Google maps might give turn by turn navigation but that is an overwhelming and largely irrelevant mess of details for most planning purposes. Eg: all I might want to know is the different flight legs and the fact that the journey takes 15-18 hours, and not all the turns and traffic lights to get from home to the airport.
I want a zoomed out picture, and to be able to fill in detail hierarchically, on demand.
Instead, one-step models give you the full high-res local structure of the graph that would have to search through (with too many states and edges).
Ha, interesting. I wasn't aware of Sutton's blog post, but if I might make a shameless plug, we demonstrated [1] exactly this problem (see section 4.4.3), and how multi-step world models (using diffusion models as the substrate) could be one potential answer.
Since then, I have come to like temporally-abstract models more and more. Rolling out in time -- either step-by-step or many steps at once -- suffers from the tyranny of the specific. For long horizon planning with agents, I care (often only approximately) about where I can end up, and seldom about exactly when I end up there. Successor features, GVFs, Forward-Backward representations, and the like seem like they have an elegant approach for structuring thinking at a "high level", instead of generating exponentially large search trees by rolling out microscopic world models.
It'll get paid from revenue, not by redirecting employee salaries. All that AI+compute is literally what customers pay Anthropic for.
Big AI labs are not software companies where payroll dominates expenses. They're capex-heavy industrial entities; it just so happens that the "machines" (whose output they sell) are nominally the same category as the devices that their knowledge worker employees use on their desks.
I doubt they're the first solution to use coordinate based editing, or even the best one right now.
Eg: Check out hash-anchored editing. The first place where I recall seeing this was the oh-my-pi coding agent, but I wouldn't be surprised if the idea originated earlier/elsewhere.
I wonder whether CRDTs could be a good solution for multiple agents editing the same codebase in parallel.
The analysis and synthesis approaches to understanding systems have respectively been the driving forces for two major breakthroughs in 20th century physics: reductionist and emergent phenomena. Reductionism aims to understand a system by reducing it to component parts and assuming that the composition is simple. This attitude drives particle physics. On the other hand, one of the most important themes from condensed matter physics has been how "more is different" and collective phenomena can induce emergent behaviour which is very different from the behaviour of the constituent parts. In this perspective, the precise constituent parts don't really matter too much -- many substrates which like completely different to analysis can end up looking very similar in synthesis. This is the principle behind universality classes in critical phenomena. This patter of thought should also be familiar to folks who advocate for a "systems perspective".
In the language of differential vs integral calculus, you can have perfectly well behaved and physical functions whose derivatives can completely miss the global behaviour of the function i.e. smooth but not analytic eg exp(-1/x) at x=0. Funnily enough, this is exactly the form of the action taken by quantum mechanics and an argument for how the classical limit irrecoverably ignores the physics of quantum systems.
> Without the price cut, Deepseek V4-pro tokens would have cost more than resold Opus 4.8 tokens.
You mean it's functionally as if American tokens are being price dumped in China and Chinese model providers are being forced to compete with that and innovate? So many delicious layers of irony, lol :-P
I don't understand the justification for local hardware with cost as the motivation. The same (or bigger/better) open weights models can served by third parties at much higher resource utilisation, and will therefore be much cheaper!?
Especially because the world is likely to persist, at least for a while, in state where computing hardware demand drastically exceeds supply resulting in high prices for hardware. So why wouldn't you want to max out utilisation and amortize costs, at least for typical (non sensitive) use cases.
You need a full ring of sensors to be receiving at all times. The Caltech design has a rotating element that emits once at each location on the ring. The Midjourney design instead uses the same receiver elements to also emit one at a time, in turns. But wherever the emitter fires from, all the receivers need to be listening.
If you move a single sensor around then you need to multiply the scanning time by the same amount, and repeat the same "experiment" for each measurement position.
When we measure the average experience, it's crucial what we are sampling/measuring uniformly to construct that experience.
The service provider is choosing to weight all requests uniformly, and average over requests -- some have 10s latency and some have 1s latency.
The user lives in time, and chooses to weight their time intervals equally. So a 10 second pause carries 10 times more weight for them than a 1 second pause -- because they experience it 10 times as much! So their average experience is a different weighted average.
The conceptual point is that averaging always needs a measure, and implicitly assumes one if you aren't explicit about the choice.
I did a tech deep-dive into the Midjourney tank, and this is basically the origin of that (the first author David Garrett worked at Midjourney for a while) so I have some thoughts on the tech.
A ring of devices at 60cm (70cm for Midjourney) diameter means much longer distances than even the deepest tissue imaging ultrasound is typically used for. Geometric 1/r^2 attenuation through water and exponential attenuation through tissue. Sensitivity is always the key question for any sensor. This is why they use 1MHz ultrasound (lower frequencies attenuate less) rather than say 5-10Mhz, but that also means lower resolution. Sound wavelength in water/tissue at 1MHz is roughly 1.5mm. So in pursuit of sensitivity, good resolution becomes a challenge.
What sensors could we line up around the tank? This work has gone to great lengths to build custom transducers for receiving the ultrasound signal, but I couldn't easily find a direct characterization of sensitivity (pardon me if I missed it; I'd love to know more) but I don't think they are claiming a breakthrough in sensitivity -- only a new (tank) system architecture. AFAIK, all available technology, be it Butterfly/Olympus/whatever have a fundamental tradeoff in cost and sensitivity, because they all pick different operating points on the Pareto curve set by the underlying technology. Butterfly is not as sensitive as the top-grade devices, but the top-grade hospital ultrasound machines are way more expensive and bulky (roughly 150k USD & 150Kg). It would be a real challenge to fit dozens of them around the tank! Let alone pay for them, if one wants to deploy at scale.
If cost and bulk were not a constraint, we could definitely use many hospital-grade sensors and get high-fidelity images. The problem is what happens if the sensitivity is not quite enough. Look at the online discourse from doctors about how low-sensitive full-body scans are asking for trouble from over-diagnosis (incidentalomas) and iatrogenic side-effects. Low sensitivity full-body scan is basically the diagnostic version of p-hacking. Just like the US air force found that there is no "fully average" pilot, every body scanned will always have some anomaly or the other. And most of them never need to be clinically intervened on. That's not quite the fault of the device, but that's still a massive gap in the deployment strategy and it will require retraining doctors and restructuring large parts of the existing healthcare system.
It's easier to just invent technology and build higher-fidelity sensors. This device is mostly just an assemblage of available sensors, we need innovation that produces sensors which are smaller + inexpensive + less power hungry. It's not just a question of assembling it into different form factors or scaling production, or nudging people into using these by frequenting spas. AFAICT that's barely getting any discussion in the chatter around the Midjourney launch.
Lastly, it's really important that this doesn't require a trained sonographer. That's a real scaling bottleneck if we want to enable exponentially more usage of scanning. (Not taking a dig at the paper authors, but only at the online discourse) If we really wanted to scale this to population level use... the tank form-factor is good, but guess what's easier: having smaller devices one could just drop into a bathtub at home :-)
When aiming for 100k tok/s, you would still have CUDA overheads (on the order of microseconds) -- which might become the bottleneck, even if you do everything else right with the inference architecture. How are you planning to overcome that?
EDIT: Oh, on second read, do you mean you're running the model on an FPGA?
My apologies... I was responding to the above comment / ranting about the general trend and got carried away. Wasn't directed at specifically at your post.
I love your second graph; hope the trend catches on as the main graph, instead of the model-wise bar graph that seems to be popular.
I don't specifically care about Claude -vs- GPT, but comparing models at different amounts of test time compute is a gaping hole. It also means that any unreasonably-expensive token guzzling white-elephant model can top all the benchmarks and still be useless.
What we actually have is like a scaling law for test time compute, so it's silly to focus on specific Y values that someone benchmarked (at whatever default X values). Instead, characterize the slope or power of the scaling law, or just plot the damn curve for each model -vs- number of tokens or cost or something!
email: siva dot rk dot sw at Google's email service
(Opinions my own; not representing any entity)