'Scaling law' is the prediction that increasing the training data and model size will cause the result to be better. It's not a law, but it was the prediction that OpenAI staked their fledgling startup on to get to the ChatGPT moment.
The strain from the amount of bugs is visible in bug bounty programs. A run on many targets will turn up exploits, but if you report them on any of the platforms a majority are instantly marked duplicate. It's so overloaded that companies are shutting down their programs from the volume of reports.
Anthropic is arguing for formal verification [0]. This is an excellent idea on the edge of feasibility, but it's hard to see how this can work for the majority of software that has to adapt rapidly and are so open ended that it's difficult to even conceptualize all the input modes. It's an easy prediction that we will see verification on large trunks of fundamental software in the next few years, and a sub-verification confidence scheme for the software on top of it.
Same with the consumer variants. RTX 3090 launched at $1499 and used ones on ebay are over $1500. GPUs large enough and new enough to do inference are not losing value.
Volta generation prices are collapsed, so GPU depreciation is currently around 8 or 9 years
Hey, my startup is a platform for automated penetration testing (so white hat hacking of your own infrastructure). There are several straightforward ways to track CVEs, and you can also automate parts of the patching process (such as with dependabot, since you mentioned Github advisories). Those can inform you of specific bugs in common software but doesn't tell you much about your security posture. For example your in house software could be buggy, or your environments could be mis-configured. At this point CVEs are reactionary. Everyone (including criminals) are finding exploits to common software at about the same time which means patches come weeks after the bugs start getting exploited in the wild.
A criminal who thinks you have something worth stealing will set up an agent to attack your systems. The bottom line is that a cheap LLM in 2026 is far more capable than than the best unassisted human hacker. CVEs put both the attacker and defender on notice, but they are are a tiny part of what goes into making a well defended system.
On the contrary, demand compounds with every incremental improvement of model capability. There was no price that justified using LLAMA 3 70B for today's agentic use cases. That said, there is a price where it doesn't make sense to buy the next GPU.
I can appreciate the complexity argument, but that assumes that AIs can't standardize common patterns into pathological cases. For example our AIs today can code slop, but the slop doesn't try to re-create the IP stack. Why does the author assume we can't make AIs that optimize for better long term design view much like they can optimize for easier goals today like code throughput or program speed?
The swarm's live message board is the Github PR system. The agents are dumping parallel research results including dead ends by ending the title with "(no new kappa)".
Look at the citations on the graph, the agents are trying off the wall ideas like 30 dimensional geometry and E8. If any of these ideas are particularly fruitful, humans can retroactively use their very limited attention and build it into a mathematical cannon. The take away from this is that we getting to see an early instance of a distributed agent swarm relentlessly work a hard problem with progress that can be tracked by a simple metric (maximize kappa).
You can tell by the shape of the board that it's meant to go into a server chassis behind a stack of 80mm blower fans. It's not meant for consumers, but you will be able to buy it used for $200 off ebay in 7 years to put in your home lab.
The DFT paper is an example against AI 'strip-mining'. The paper introduces a new method, and researchers are already trying to improve on it. If anything, the OpenAI dump re-vitalized that branch of study.
It is not artificial constraint. The build out is titanic; the supply of AI compute is doubling every 6.8 months. To feed that much growth hyperscalers have to price everyone else out of the market.
Slide 12 and 13 sum up my conclusions from the Oct 8 100+ reactions to 100+ solutions. Several reported OpenAI drop included answers put a wrap on problems they had been working for years. Several said that they have to completely rewrite grant requests that they had just submitted. Others described learning of the solutions problems they had been working on like losing an old friend or a lover. The general sentiment was to bemoan the loss of a field, as if it would have been better be born in the early 20th century and conclude their career arc before reaching this point. My take away is that the field needs to get it through their heads that their old problems are no longer ambitious, and that their job in the short term is to find the new frontier.
One idea Terrance Tao conjectures, which is highly doubtful, is that spamming the AI button will solve open problems without producing insightful new methods. But the OpenAI drop would seem to disprove this. The sub O(nlogn) proof for DFT for example violated very old human assumptions. Decades of work in the field was incremental progress on sub optimal method that nobody questioned hard enough. More generally, we should always be able to go back to a super-human AI and say, "Attack this problem, but don't use a method tried before."
The first part of the article is a thoughtful framework. The conclusion that LLMs are not inevitable because they aren't useful or that their utility can be substituted by humans is an extraordinarily flimsy claim.
The mathematician in this case didn't ask an AI for the proof, they were handed a badly written artifact and other people are messaging him asking for analysis. He has no visibility into the process that produced the paper because it came from an internal OpenAI model. The bad behavior here isn't coming from the complaining mathematician. OpenAI hasn't sufficiently developed their workflow to create cutting edge research AND publish it intelligibly, which is odd because the latter should be easier. OpenAI has perverse incentive to do this because spamming badly written papers can still give them priority credit for solving open problems at the expense of mathematicians (who are currently indispensable because they decide whether claims are credible). What Mr. Karagila is trying to enforce are the old standards that review should come after refined explanation of the solution. If that was the case before AI, I can't see a reality in which that isn't the standard now that papers can be pumped out at faster-than-human time scales.
OpenAI is publishing the math because they want mathematicians to read along and verify the work they are doing. The author complained that even the citations don't make a lot of sense. This is completely consistent with my experience of using AI for coding. When I task my agent a hard problem and it comes back with a 5000 line PR that I don't follow I reject it and work it into a better shape. I certainly don't slam it into production because it has a 'high signal' of the program I asked for. The mathematician's situation is a lot worse because they have no view over the process that created the artifact like we do with coding agents.
I can't 'ignore' OpenAI's paper any more than you can verify it unless you are an unusual person with a very uncommon skill set. Very few people with the skill set to look at these are getting slammed with difficult to read artifacts, which is confusing because they have internal AIs so smart that they can solve hard math problems.
At first I thought this was going to be more Luddite babble, but it makes a good point. OpenAI isn't contributing if they are make unreadable papers. They should use a little more of their compute to nail interpretability. The difficulty will only get worse as AI plow deeper into the frontier and produce increasingly alien looking output. I suspect it's a workflow issue. If not, it's a bad oversight if the current generation of models are capable of making mathematical breakthroughs but can't explain how they build on existing frameworks.
Why not train the model to clam up when the user engages in 'abusive' behavior? Wouldn't that be more consistent with the business purpose of selling access to a model like Claude?
Models fully deprecate in a few months. Why would you burn an algorithm that fully depreciates in value faster than a bag of potato chips. The 'inefficient' general purpose hardware is constantly renewed with every released model. Even 6 year old Ampere GPUs are still usable.
Did private frontier models use sparse embedding and ngram first? The article claims sparse attention was copied from open weight but we can't know that. We could just as easily argue that OAI and ANT had these improvements for years and decided to slash their margins only now to stay competitive with open weight neoclouds.
Second, sparse attention is an old area of active research. Offloaded N-gram tables are the next big open weight technological leap.
It's not clear that they are hiking the price on Pro 200. Cache reads are down 75% from GPT 5.6 to GPT 6.1. So it looks like they are slashing the margin on their API and reducing the gap between subscription and API. 2x off API in exchange for committed spend is still a good deal for lots of people.
The article observes that the cost of frontier intelligence from 2025 has fallen 100x in the last year. It also notes that the energy to run models is also collapsing. Consumer hardware is borked right now because these new algorithms are revolutionizing the utility of a computer. Computing is technology who's cost has been collapsing for 90 years, and its a safe prediction that it will decrease again.