These are relatively contained private credit markets though. We’re not looking at anything 2009 level. For scale, total US mortgage debt peaked at $9.3T ahead of the subprime mortgage crisis, 73% of GDP at the time. We’re talking here about ~5% of GDP.
We were around ~$11 trillion in debt in 2009. Now we are ~$40 trillion in debt. The FED balance sheet was less than $1 trillion before the crisis in 2008, now it is around ~$7 trillion. Our debt to GDP ratio was ~64% in 2008 and is now at ~120%. Put simply, our ability to absorb any sort of financial shock by taking on massive amounts of (more) debt (which is what we did in 2008) is not remotely similar to what it was during the last crisis. This would be true even if we weren't in the midst of a highly inflationary environment due largely, but not solely, to the ongoing energy crisis being worsened by the wars in Iran and Ukraine.
But this assumes Chinese models will not achieve token cost optimization. Intelligence needs are fairly flat for many tasks, and the Chinese models have caught up on this front. Next they achieve greater token cost efficiency and we don’t need OpenAI.
That the author doesn't acknowledge the relentless R&D efforts DeepSeek has been plowing into optimization, and giving a default win to OpenAI/Anthropic on the supposition that they've been serving models for longer is a black mark against the article.
I appreciate the transparency in explicitly stating their motivation for writing the article (a response to what the author saw as an overreaction to Chinese models), but I feel the article goes too far the other direction, with multiple unsupported leaps of logic, and overstating the stickiness of AI client products.
The model that is most optimized around token cost is, in fact, Chinese. DeepSeek is astoundingly cheap by default, but if you use it from Reasonix (the harness optimized around its cache), it becomes even cheaper.
This also comes with significant capability reduction. deepseek-v4-flash is very good in the < 250K range, then degrades between 250-500K, and is practically unusable after 500K.
[edit]
This is my observation from using it without an specific context engineering to optimize for Deepseek's cache compression and sparse attention mechanisms. I am pretty sure that if you specifically structure your context to align to the cache compression boundaries you can significantly improve performance in the full 1M context, but there is not much reason to do this, because if you design your outer loop to work with shorter contexts that solution is portable and more efficient, so I haven't bothered with a optimizing for DS at this point.
Btw - assuming NeurIPS reviews aren’t garbage tomorrow, I’ll have a paper out which claims that most long context problems in models are really sampling problems in disguise
Switch to a modern sampler like min_p or ideally a better one like top-n-sigma (it’s in llamacpp) and your “my model gets stupid at long context problems” will basically go away.
Unfortunately this fact is still not well appreciated yet despite nearly every modern sampling technique getting an oral wherever they get presented. Min-K just got an oral at ACL 2026, for a hyper recent example of this. There’s a reason they keep getting orals.
The field massively ignored sampling for mostly safety reasons and now the whole field incorrectly believes long context doesn’t work on small models. Long context is an out-of-distribution problem. Your sampler configured properly keeps you in distribution.
Oh and this is doubly true for quantized models. I run my qwen 3.6 27b with 4bit quants from unsloth and get excellent performance because my sampler stack is good and not the garbage that is top_p and top_k. Also, yes, you need to ignore the trash recommended sampler settings from the Chinese labs (they’re wrong/bad).
Interesting. I am not familiar with model internals at this level because I have only been working at the application layer so far, but will definitely research this further. When you get the paper published would appreciate if you can drop a comment with the link so I can read it.
I've used RLAIF to build out heuristic based non-LLM models for various decision systems and achieved like, 95% F1 on certain projects. We're in a place where models can be used to fine tune a lot of stuff via loops.
LLM architectures need to fundamentally change or inference needs to be used in constrained trusted environments. Nothing surprising here. Filtering and sanitizing, relying on tags around input strings that can be intercepted and replayed is like, childs play security theatre. As long as prompts accept abitrary user input nothing is changing here. Non-deterministic security is never going to be acceptable.
I'm quite certain that Google's AI services are likely the most used in the world right now by virtue of having the widest distribution. It's in the search box. It's on your Android phone. Just because they aren't the preferred coding or research agent does not mean they are losing - that's a pretty small slice.
Yeah this seems true. Claude Code are famously dubbed as best AI coding agent, but google doesn't care about that niche I guess. Somehow, I still rely on google search as they have diversified it.
If you ask questions, it will enable "AI overview" , but if we search about particular object/platform like "Google stock" or "bbc news", it will give the old classic search experience and we woulnd't need to swallow "AI overview" pill in that case.
Would love to see this benchmark tested on more perceivably LLM friendly frameworks/ORM (e.g. is NestJS or Drizzle / Kysely more performant than their choice of Sequelize) and more frontier model vs just GPT 5.2.
Anyone read whether these tests include any validation loops? What happens if the models get back test failures, for instance? Understanding how many turns to hit full passing behavior suite would also be interesting. Great methodology in the study though.
This is fairly standard practice for device fingerprinting. LI is probably using this to protect its platform from scraping etc, and extension lists have sufficient enough entropy to help identify users and form a useful component of a fingerprint.
Its already pretty easy to oneshot an extension aiding scraping and LI can do nothing about it. I've seen people build and install a local chrome extension in a couple of days and have an AI inject itself into devtools and scrape pretty much any website. And that was a few months ago. I don't think there is an easy way to defend against such things anymore. Its a matter of time that defensive programming measures like this become useless.
The worst has been the post-covid assignment of seating and QR code driven ordering in bars. So few opportunities to mingle. I miss standing in bars, talking to bartenders, chatting with random patrons. This has recovered much better in large cities but I find that restaurants and bars in US suburban environments are deeply impersonal now. It’s no wonder singles are stuck meeting partners on apps with so little unstructured social opportunities left. Not to mention no one is going to bars anymore anyway.
I'm still stuck on superpowers. Can't seem to get better plans out of native claude planning - superpowers ensures I have a reviewed design that actually matches my mental model. Typical claude planning doesn't confirm assumptions sufficiently for my weak brain dumps/poorly spec'd tickets.
The thing with microbes is not if they can grow in a place it's whether they can get there first.
Beer is basically knocking out natural bacteria and trying to get yeast growing before the bacteria can turn it into cleaning supplies. The alcohol is kind a there because it kills bacteria.
So for instance I put winecaps (Stropharia rugosannulata) into wood chips that had already been exposed to the elements for six months, and ended up with more than I could possibly eat.
Meanwhile oyster or shiitake mushrooms want a fresh log, cut with a sterilized blade, and cross your fingers and hope. I haven't even tried because I've watched people who know way more than me about mushrooms, fail.
I think I have some logs that might have lions mane in them, but they're fighting the turkey tail that was already in my local environment and also on the property of the person who donated the logs.
