Artificial Intelligence

[ajlomb1011]

21 hours ago, beerbaron said:

AMAZING. This is a model on par with top end frontiers from december 2025. 1/15th of the price of opus. Imagine if China pulls out a better coding frontier model just before the 3 big IPOs, that would cratter valuations.

 

How many over here know if there company allows dev. with chinese models? 

 

I

That’s why Anthropic and OpenAI want to IPO ASAP. There isn’t a moat in LLMs, and there never will be. 

[LC]

On 5/23/2026 at 4:11 PM, beerbaron said:

How many over here know if there company allows dev. with chinese models? 

 

Large FI: we don't

 

I'm surprised there isn't a US clone of deepseek yet. It would certainly take some share as most use cases don't need frontier LLMs and so price becomes a more competitive factor. IMO I agree as @ajlomb1011 says ultimately LLMs will move towards commoditization like other general compute.

 

Might take a few years for (1) a US variant and (2) corporate america to fully embed AI and fire 50% of their workforce, fully relying on LLMs and then pressing that cost line.

[Milu]

I feel like for stuff like coding where AI is really starting to show great promise  models that are at the cutting edge will always be in high demand. So as long as tech new Claude or Open AI coding models are staying 3 months ahead of the cheaper open source models they will have loads of demand. Even though 5 months seems like a short time, the difference is quality in answers and coding from a December 2025 model and a May 2026 model is massive. Obviously if they can’t maintain this edge then it’s a different story. 

[Spekulatius]

18 minutes ago, Milu said:

I feel like for stuff like coding where AI is really starting to show great promise  models that are at the cutting edge will always be in high demand. So as long as tech new Claude or Open AI coding models are staying 3 months ahead of the cheaper open source models they will have loads of demand. Even though 5 months seems like a short time, the difference is quality in answers and coding from a December 2025 model and a May 2026 model is massive. Obviously if they can’t maintain this edge then it’s a different story. 

3 month may be a large lead now,  it it’s likely not a likely long lead in 2 years, as eventually progress will wane.

 

Most large companies will not use Chinese AI but scrappy upstarts will. I also think the situation is differnt in other countries, especially those that need to make every $ count.

Edited May 24 by Spekulatius

[Milu]

Just now, Spekulatius said:

3 month lay be a large lead now,  it it’s likely not a long lead in 2 years.

 

Most large companies will not use Chinese AI but scrappy upstarts will. I also think the situation is differnt in other countries, especially those that need to make every $ count.

I feel like we are just at the beginning with these models and some march towards AGI and ultimately ASI over the next couple of decades. Today’s cutting edge models will look like toys compared to next years cutting edge models and this will continue for a long while yet. Only the big frontier models will have the capital and chip access to afford to stay at the bleeding edge, then open source models will constantly be behind due to this fact. Maybe the gap will narrow slightly, maybe it will get larger. And as long as this is the case, businesses who want the biggest edge on their completion will be more than willing to pay up for the best.

[frommi]

9 hours ago, Milu said:

I feel like we are just at the beginning with these models and some march towards AGI and ultimately ASI over the next couple of decades. Today’s cutting edge models will look like toys compared to next years cutting edge models and this will continue for a long while yet. Only the big frontier models will have the capital and chip access to afford to stay at the bleeding edge, then open source models will constantly be behind due to this fact. Maybe the gap will narrow slightly, maybe it will get larger. And as long as this is the case, businesses who want the biggest edge on their completion will be more than willing to pay up for the best.

Do you have data and facts about this? Because from what i see on benchmarks is clear diminishing returns. Models don't really get better and thats why these open source models have catched up. In fact they are so good that they already have a marketshare of 30% of global ai usage.

[ajlomb1011]

6 hours ago, frommi said:

Do you have data and facts about this? Because from what i see on benchmarks is clear diminishing returns. Models don't really get better and thats why these open source models have catched up. In fact they are so good that they already have a marketshare of 30% of global ai usage.

Open source models being more widely adopted should be a warning to anybody that thinks this capex explosion is sustainable.

 

Running open source models locally is where the future is headed, especially if chips get more power efficient with AI. This is why I think the current data center narrative that the entire market is being held up by is very dangerous. 

 

Also anyone claiming “AGI” is coming is just talking about of their ass. It’s impossible for AGI to appear out of thin air from throwing more compute at LLMs.

[Milu]

6 hours ago, frommi said:

Do you have data and facts about this? Because from what i see on benchmarks is clear diminishing returns. Models don't really get better and thats why these open source models have catched up. In fact they are so good that they already have a marketshare of 30% of global ai usage.

There tends to be an accordion type narrowing and expansion in that the frontier model firms release a new model and gap widens, then open source catches up and narrows it somewhat, then new breakthrough or development at frontier pushes next model far ahead again.
 

I would expect open source models to continue to maintain or grow market share as firms mainly utilise a hybrid mix of frontier models and open source depending on the request. For example new requests would feed into the frontier model as the brain which would then determine whether to feed it out to an open source model for straightforward requirements or the most optimal frontier model for more complex issues. 
 

Gaps in reasoning (between frontier and open source) of even just 5% translates to millions of dollars of alpha for businesses vs their competitors. Some examples of this would be in the area of drug discovery and algorithmic trading where being 6 month behind cutting edge is a no starter.

 

While I know many are predicting/hoping that all this capex is going to be wasted and that the mag 7 and Anthropic/Open AI/Spacex are going to collapse like a tower of cards, I fear they might be disappointed. Doesn’t mean their won’t be lots of volatility and various drawdowns in these stocks over coming years.

[Spekulatius]

16 hours ago, Milu said:

I feel like we are just at the beginning with these models and some march towards AGI and ultimately ASI over the next couple of decades. Today’s cutting edge models will look like toys compared to next years cutting edge models and this will continue for a long while yet. Only the big frontier models will have the capital and chip access to afford to stay at the bleeding edge, then open source models will constantly be behind due to this fact. Maybe the gap will narrow slightly, maybe it will get larger. And as long as this is the case, businesses who want the biggest edge on their completion will be more than willing to pay up for the best.

I think it’s more likely that the current crop of models matures and progress slows. My guess is that new models using a different foundation that more mimic reasoning not guessing are needed to achieve AGI which is a far way off. AGI means that models can deal with datasets and domains they have not been trained on which is currently not the case.

Edited May 25 by Spekulatius

[frommi]

1 hour ago, Milu said:

Gaps in reasoning (between frontier and open source) of even just 5% translates to millions of dollars of alpha for businesses vs their competitors. Some examples of this would be in the area of drug discovery and algorithmic trading where being 6 month behind cutting edge is a no starter.

We have algorithmic trading since 30 years. LLM's are really not good at that. For drug discovery the frontier models are also not used, these are specially trained models. And thats probably the future, lots of specialized models. But without data and incremental better data, these models don't get better. And don't forget that 5% better isn't really measurable outside of benchmarks (which are all gamed because the results feed back into the next model, essentially curve fitting the model to the benchmark), with such a tiny difference cost is a bigger factor.

Edited May 25 by frommi

[Spekulatius]

I think parettos law applies here. If you get 80-90% of the performance for 20% of the cost , then it probably makes economic sense to run a cheaper model, which can be run be run cheaper hardware as an additional benefit, train this model using your own data (the fact that ChatGPT can answer questions about the 2nd Punic war and your model can’t is not relevant ) and you have a much more cost efficient and leaner solution that works for your business case.

[Milu]

It will be interesting to see how this plays out. A lot of capex being spent on the hope that frontier models continue to improve and businesses will see value and spend. And then a lot of good open source models trying to keep up. Most people alway fall into two binary camps, one is that the frontier models will continue to add value that justify the crazy spend, and the other is that these companies will all go bust as open source catches up. In reality, like with most other things some middle ground will likely prevail where maybe some over spend occurs but these frontier models continue to show value while also being back up by top quality open source solutions. And businesses will use both.

[nsx5200]

Like all previous "breakthroughs"(oil/electricity/broadcast medium/internet), it will follow the S-curve instead of a linear or exponential curve.

 

There are many signs that we're at the start of the levelling off of the S-curve.  We're starting to see more AI generated contents, which is used to train future generation models.  We know that performance of models that train on AI generated content degrades, and we also know that they're starting to run out of human-generated contents to train future models on.  We're starting to see the many constraints, many of them are physical now, of what it takes to build out an AI-centric future.

 

At some point, incremental model improvement will slow down.  Any future improvement will be grindier.  Current AI will at some point be commodity similar to previously mentioned breakthroughs.  IMHO, China/CCP understand correctly that AI is just a new tool, and they're using this new tool to apply it when it makes sense.

[MungerWunger]

 

Someone made a Chinese love movie starring the Anthropic CEO using a Chinese AI model. Incredible ...

 

Edited May 25 by MungerWunger

[Spekulatius]

3 hours ago, Milu said:

It will be interesting to see how this plays out. A lot of capex being spent on the hope that frontier models continue to improve and businesses will see value and spend. And then a lot of good open source models trying to keep up. Most people alway fall into two binary camps, one is that the frontier models will continue to add value that justify the crazy spend, and the other is that these companies will all go bust as open source catches up. In reality, like with most other things some middle ground will likely prevail where maybe some over spend occurs but these frontier models continue to show value while also being back up by top quality open source solutions. And businesses will use both.

I don’t think the middle ground scenario is baked into the valuations of memory stocks, semis, AI Capex winners and whatever gets IPO’d as AI stocks though. The market has become one big AI trade.

[roundball100]

7 hours ago, ajlomb1011 said:

Open source models being more widely adopted should be a warning to anybody that thinks this capex explosion is sustainable.

 

Running open source models locally is where the future is headed, especially if chips get more power efficient with AI. This is why I think the current data center narrative that the entire market is being held up by is very dangerous. 

 

Also anyone claiming “AGI” is coming is just talking about of their ass. It’s impossible for AGI to appear out of thin air from throwing more compute at LLMs.

Running open source models locally also makes a whole lot more sense from a privacy perspective (avoiding leaking the crown jewels, proprietary and private data, to an untrusted  and often outsourced third party) cloud service. 

[Milu]

2 hours ago, Spekulatius said:

I don’t think the middle ground scenario is baked into the valuations of memory stocks, semis, AI Capex winners and whatever gets IPO’d as AI stocks though. The market has become one big AI trade.

Ya probably not. I haven’t been buying anything in that space. My existing holdings of meta, google, amazon etc are obviously giving me some exposure to the AI trade but I’ve owned them for a long time. Well before AI became a big buzzword in the markets. I’m not planning to sell these if they get a bit overvalued due to an AI bubble. Semis and other AI related stuff I don’t really have any skills at evaluating so any new investments I’ve made have been mostly outside of any AI impacted industries (Dominos, Ferrari, Hermes, small amount of Fairfax)

[MungerWunger]

https://www.businessinsider.com/uber-coo-andrew-macdonald-ai-token-spending-harder-justify-2026-5

 

"In a Rapid Response interview released on Saturday, Uber's operations chief, Andrew Macdonald, said it was becoming harder to justify AI costs within the company."

 

"That link is not there yet, right?" he said. "I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25% more useful consumer features.'"

[Spekulatius]

This is a common mistake with AI. Try it yourself with your favorite chatbot. We are not at AGI yet.

 

 

 

I think the LLM “think” you are asking for how many “r’s” are in the word strawberry ignoring your precise questioning.

Edited Saturday at 12:51 PM by Spekulatius

[gfp]

meanwhile...

https://www.wsj.com/tech/ai/ai-math-solves-erdos-problem-openai-c4029e84?st=ZZ2yUC&reflink=desktopwebshare_permalink

[beerbaron]

1 hour ago, Spekulatius said:

This is a common mistake with AI. Try it yourself with your favorite chatbot. We are not at AGI yet.

 

 

 

I think the LLM “think” you are asking for how many “r’s” are in the word strawberry ignoring your precise questioning.

I think these examples are mistaking the trees from the forest. This is just a function on how LLM operates. Does not show lack of intelligence but rather how it see things (token).

[nsx5200]

 

We're going to see a whole new suite of attacks and volatilities from these agentic trading.  Who said value investing is dead?

[Spekulatius]

On 5/30/2026 at 10:17 AM, gfp said:

meanwhile...

https://www.wsj.com/tech/ai/ai-math-solves-erdos-problem-openai-c4029e84?st=ZZ2yUC&reflink=desktopwebshare_permalink

This is amazing and the proof seems quite elegant. The Ai used a completely different toolbox (Number theory) to come to a solution for this geometrical problem.

Edited Sunday at 03:20 PM by Spekulatius

[Paarslaars]

On 5/30/2026 at 2:48 PM, Spekulatius said:

This is a common mistake with AI. Try it yourself with your favorite chatbot. We are not at AGI yet.

 

 

 

I think the LLM “think” you are asking for how many “r’s” are in the word strawberry ignoring your precise questioning.

 

Copilot gets it right though.