I would say “for now” and that Anthropic’s inability to execute (3.7 was only released in response to competition from DeepSeek) raises a lot of questions about their future. While I agree that DeepSeek is still behind, Anthropic also does not have the same caliber of ambitious talent that DeepSeek has and has become a magnet for safety-obsessed cranks, whom (in total) act as a drag on the company. That Dario has also been engaging in dull-minded saber rattling about China does not speak well about their ability to deliver (vs. posture) in the future.
China’s strategy is to prevent any one bloc from achieving dominance and cutting off the others, while being the sole locus for the killer combination of industrial capacity + advanced research.
Anyone doing this work should be building on top of the OpenAlex API and its Unpaywall integration. This allows you to traverse the world’s largest research graph and access full text, where it’s available in any publicly available form.
The problem isn’t the prestige it’s that prestigious institutions in America don’t produce high-quality talent. They’re instead mostly corrupt credentialing mills for the rich and well-connected. From what I understand, DeepSeek also only hires from the best universities in China, but “best” actually means something relative to how difficult entrance to those organizations is to achieve and their coursework.
It’s funny because I also have found myself doing this exact with R1+Sonnet 3.5 recently. Windsurf allows you to do a chat mode exchange with one model and then switch to another to implement. The reasoning models all seem pretty poorly implemented for the agentic workflows, but work well when paired with Claude.
If Buffet is fearful, the people being greedy are so beyond your means and chasing an irrational self-destructiveness that you have zero chance of seeing an upside by being aligned with them.
My favorite thing about LibreOffice is that it can open CSVs without breaking them. Pretty much everything else about it is a slow, unintuitive mess, but this one killer feature makes me always install it and something that somehow Excel does not have.
o3 mini’s date cut-off is 2023, so it’s unfortunately not gonna be useful for anything that requires knowledge of recent framework updates, which includes probably all big frontend stuff.
This is not creating art, but illustration. Art is where the meaning is inherent to the work, whereas illustration is when it is derived from something else that it accompanies (e.g. a text). There’s obviously a spectrum here, but GenAI blogpost decoration are pretty much the terminal point on the illustration part of that scale.
Interested to see what folks do with putting DeepSeek-style RL methods on top of this. The smaller Mistral models have always punched above their weight and been the best for fine-tuning.
> “people should not be surprised when they are occasionally surprised, and they should anticipate being surprised some more in the future.”
It’s mildly hilarious that you’re deriding Taleb for thinking (and investing) at a level that all of bubble-chasing investment houses are incapable of achieving. He’s made so much money precisely because this thing you’re describing as so stupidly obvious, is actually, not, stupidly obvious to people.
So context size actually helps with this, relative to how LLMs are actually deployed as applications. For example, if you look at how the “continue” option in the DeepSeek web app works for code gen, what they’re likely doing is reinserting the prior messages (in some form) to a new one to prompt further completion. The more context size a model has and can manage successfully, the better it will likely be able at generating longer code blocks.
It doesn’t call it into question- they’re not. OpenAI has been bleeding researchers since the Anthropic split (and arguably their best ones, given Claude vs GPT-4o). While Google should have all the data in the world to build the best models, they still seem organizationally incapable of leveraging it to the their advantage, as was the case with their inventing Transformers in the first place.
I mean what’s also incredible about all this cope is that it’s exactly the same David-v-Goliath story that’s been lionized in the tech scene for decades now about how the truly hungry and brilliant can form startups to take out incumbents and ride their way to billions. So, if that’s not true for DeepSeek, I guess all the people who did that in the U.S. were also secretly state-sponsored operations to like make better SAAS platforms or something?
Uh, they invented multilatent attention and since the method for creating o1 was never published, they’re the only documented example of producing a model of comparable quality. They also demonstrated massive gains to the performance of smaller models through distillation of this model/these methods, so no, not really. I know this is the internet, but we should try to not just say things.
4o is more expensive than DeepSeek-R1, so…? Even if we took your premise as true and we say they are as good as DeepSeek, this would just mean that OpenAI is wildly overcharging its users.
> It can solve that Arc-AGI benchmark thing at huge compute cost
Considering DeepSeek v3 trained for $5-6M and their R1 API pricing is 30x less than o1, I wouldn’t expect this to hold true for long. Also seems like OpenAI isn’t great at optimization.
I used to think this, but using o1 quite a bit lately has convinced me otherwise. It’s been 1-shotting the fairly non-trivial coding problems I throw at it and is good about outputting large, complete code blocks. By contrast, Claude immediately starts nagging you about hitting usage limits after a few back and forth and has some kind of hack in place to start abbreviating code when conversations get too long,
even when explicitly instructed to do otherwise. I would imagine that Anthropic can produce a good test time compute model as well, but until they have something publicly available, OpenAI has stolen back the lead.
The same way you check performance for any problem like this: by creating one or more manually-labeled test datasets, randomly sampled from the target data and looking at the resulting precision, recall, f-scores etc. LLMs change pretty much nothing about evaluation for most NLP tasks.
>the logical next step is for payers to vertically integrate and bring more care in house where they can better control cost and quality.
…except you skipped over the part where UHC billed at higher rates for the clinics they own so they could screw customers with premium hikes, take a bigger percentage of of the inflated bills, as well as profit from whatever costs they didn’t cover and people were forced to pay.