Artificial Intelligence

[Milu]

7 hours ago, nsx5200 said:

Just curious, does changing the depreciation schedule from 4-6 year really change your investment thesis that much?

Nope, which is why I continue to hold many of these stocks in a big way and believe that mag 7 will continue to grow ever larger over coming decade and all the fear about overspend and comments about correct depreciation rates are mostly pointless. I only posted the video to help some of the bears see things better. 

[Spekulatius]

13 hours ago, Milu said:

Would encourage people who keep highlighting the depreciation estimates and whether 4 years, 5 years, 6 years is correct etc to listen to first 10 minutes of this.

 

Pay particular attention to where they highlight how google is still running their 10 year old TPU chips at 100%, so if anything their could be a case that many of these companies are possibly still being conservative about the depreciation estimates. 

 

These chips may still be marginally profitable and covering their cost of running (electricity) , especially since there is no depreciation any more. @beerbaron has a good point about technical obsolescence once the shortage of GPU capacity abides. If you don’t believe that NVDA will rapidly improve their GPU, then why is it worth $4.6T. Everyone can just run their 10 year old fully depreciated GPU‘s instead of buying new ones.  Both cannot be true at the same time.

 

 

[Milu]

5 hours ago, Spekulatius said:

These chips may still be marginally profitable and covering their cost of running (electricity) , especially since there is no depreciation any more. @beerbaron has a good point about technical obsolescence once the shortage of GPU capacity abides. If you don’t believe that NVDA will rapidly improve their GPU, then why is it worth $4.6T. Everyone can just run their 10 year old fully depreciated GPU‘s instead of buying new ones.  Both cannot be true at the same time.

 

 

Sure but can’t companies do both, use cutting edge nvidia chips for the tasks that require them, for example training runs for the models, and then use older less cutting edge chips for less complicated tasks. Overall it seems that these hyperscalers are getting a lot longer out of their hardware than was previously thought and they are adjusting their accounting to reflect this. It’s not some co-ordinated ploy to juice earnings. 

[Spekulatius]

This podcast from Galloway is preset good. the ghost is that we create ten billion $ in revenue with trillion $ investments right now. OpenAI is the prime example making $13 billion in ARR while talking about $ trillions plus in investments.

 

[LC]

I think the value prop is more about expense reduction vs. topline growth, but I think it's a fools errand to guess how this technology will be ultimately used. 

In other words, we could all be wrong, or we could all be right, or a mix of both. 

 

Also:

https://gizmodo.com/yann-lecun-world-models-2000685265

 

 

Edited November 16, 2025 by LC

[Libs]

13 hours ago, Spekulatius said:

This podcast from Galloway is preset good. the ghost is that we create ten billion $ in revenue with trillion $ investments right now. OpenAI is the prime example making $13 billion in ARR while talking about $ trillions plus in investments.

 

 

 

At 16:00, check out Altman getting defensive when asked a simple question of how the hell Open AI can commit to trillions in spending with current revenue of $13B.

 

Classic

 

[Spekulatius]

57 minutes ago, Libs said:

 

 

At 16:00, check out Altman getting defensive when asked a simple question of how the hell Open AI can commit to trillions in spending with current revenue of $13B.

 

Classic

 

His answer is basically  - “Our stock goes up.”
Generating $13B in revenue with $1.3T in investments (paraphrasing a bit here) is going to be a classic. This is basically the “you got to dance while the music is playing moment”.

 

He is not wrong either from his perspective. He is basically making OpenAI so big and entangled that it’s too large to fail and will be supported by other players in this AI spider web, if things go south.

 

Just my guess, but the real winners of the AI boom will grow from the compost left from the first investment boom we are currently experiencing in a couple of years,

.

Edited November 16, 2025 by Spekulatius

[MungerWunger]

 

 

"AI has been compared to various historical precedents: electricity, industrial revolution, etc., I think the strongest analogy is that of AI as a new computing paradigm (Software 2.0) because both are fundamentally about the automation of digital information processing.

 

If you were to forecast the impact of computing on the job market in ~1980s, the most predictive feature of a task/job you'd look at is to what extent the algorithm of it is fixed, i.e. are you just mechanically transforming information according to rote, easy to specify rules (e.g. typing, bookkeeping, human calculators, etc.)? Back then, this was the class of programs that the computing capability of that era allowed us to write (by hand, manually).

 

With AI now, we are able to write new programs that we could never hope to write by hand before. We do it by specifying objectives (e.g. classification accuracy, reward functions), and we search the program space via gradient descent to find neural networks that work well against that objective. This is my Software 2.0 blog post from a while ago. In this new programming paradigm then, the new most predictive feature to look at is verifiability. If a task/job is verifiable, then it is optimizable directly or via reinforcement learning, and a neural net can be trained to work extremely well. It's about to what extent an AI can "practice" something. The environment has to be resettable (you can start a new attempt), efficient (a lot attempts can be made), and rewardable (there is some automated process to reward any specific attempt that was made).

 

The more a task/job is verifiable, the more amenable it is to automation in the new programming paradigm. If it is not verifiable, it has to fall out from neural net magic of generalization fingers crossed, or via weaker means like imitation. This is what's driving the "jagged" frontier of progress in LLMs. Tasks that are verifiable progress rapidly, including possibly beyond the ability of top experts (e.g. math, code, amount of time spent watching videos, anything that looks like puzzles with correct answers), while many others lag by comparison (creative, strategic, tasks that combine real-world knowledge, state, context and common sense). 

 

Software 1.0 easily automates what you can specify.
Software 2.0 easily automates what you can verify."

[Spekulatius]

2 hours ago, MungerWunger said:

With AI now, we are able to write new programs that we could never hope to write by hand before. We do it by specifying objectives (e.g. classification accuracy, reward functions), and we search the program space

I don’t think this is correct, AI basically search repositories for similar code and then sort of stitches it together. So all the code that AI gives you has been written before. It is more or less a very good search engine for existing code in GitHub that is also able to guess parts and stitches it together.

 

In my limited experience using it for coding and macros, it is very good and sort of finding what you need but you have to be careful how you put it together with other coding piece from AI or otherwise. it often messes up variables and the code pieces may have assumptions that incongruous with other code pieces it created before. it’s more of a building block generator for code than a coder itself.

 

I found it helpful but you really need to check all this stuff. I have also seen it dot r math mistake because LLM fundamental don’t understand math either. Anyways, a that my limited understanding from my experience but ai really don’t code for a living, so take this with a grain of salt.

 

Because of above, I think there is a limit how far LLM can go no matter how much computing power you throw at it and something else is needed to make further progress.

Edited November 16, 2025 by Spekulatius

[Dalal.Holdings]

10 minutes ago, Spekulatius said:

I don’t think this is correct, AI basically search repositories for similar code and then sort of stitches it together. So all the code that AI gives you has been written before. It is more or less a very good search engine for existing code in GitHub that is also able to guess parts and stitches it together.

 

In my limited experience using it for coding and macros, it is very good and sort of finding what you need but you have to be careful how you put it together with other coding piece from AI or otherwise. it often messes up variables and the code pieces may have assumptions that incongruous with other code pieces it created before. it’s more of a building block generator for code than a coder itself.

 

I found it helpful but you really need to check all this stuff. I have also seen it dot r math mistake because LLM fundamental don’t understand math either. Anyways, a that my limited understanding from my experience but ai really don’t code for a living, so take this with a grain of salt.

 

Because of above, I think there is a limit how far LLM can go no matter how much computing power you throw at it and something else is needed to make further progress.

 

The LLMs function differently than something like Wolfram Alpha which is far better at math. I think a lot of the LLMs have incorporated some version of what Wolfram has to do calculations because they used to be a lot more error prone than now, but I agree that with programming, it's very finicky and requires a good amount of human oversight

[Spekulatius]

4 minutes ago, Dalal.Holdings said:

 

The LLMs function differently than something like Wolfram Alpha which is far better at math. I think a lot of the LLMs have incorporated some version of what Wolfram has to do calculations because they used to be a lot more error prone than now, but I agree that with programming, it's very finicky and requires a good amount of human oversight

Yes,  but you can run LLM as sort of a human interface for Wolfram Alpha and then it becomes a different beast. Wolfram does math correctly. Thats where I think this is going. The LLM itself becomes a commodity tool but when you tie them up with domain specific software and/ data, you can create something very powerful that won’t become a commodity right away,

Edited November 16, 2025 by Spekulatius

[brk64311]

https://www.wsj.com/tech/ai/yann-lecun-ai-meta-0058b13c?gaa_at=eafs&gaa_n=AWEtsqepWc-uIbbtEU-VjQIIMqbYykI6pS5bEAASL-_HgnE-ManIKlHZlWRKa3al79k%3D&gaa_ts=691a7b61&gaa_sig=CKqJL6gkWJdu3O5Dod8XYeXcRnsjhmbLoCI_O_JoWe5L9ev2qxBCku-YerNn4tn2tSWWjWsMGhtBg73it413sw%3D%3D

 

He’s Been Right About AI for 40 Years. Now He Thinks Everyone Is Wrong.

....

"Meta Chief Executive Mark Zuckerberg has been pouring countless billions into the pursuit of what he calls “superintelligence,” hiring an army of top researchers tasked with developing its large language model, Llama, into something that can outperform ChatGPT and Google’s Gemini. 

LeCun, by his choice, has taken a different direction. He has been telling anyone who asks that he thinks large language models, or LLMs, are a dead end in the pursuit of computers that can truly outthink humans. He’s fond of comparing the current state-of-the-art models to the mind of a cat—and he believes the cat to be smarter. Several years ago, he stepped back from managing his AI division at Meta, called FAIR, in favor of a role as an individual contributor doing long-term research. 

“I’ve been not making friends in various corners of Silicon Valley, including at Meta, saying that within three to five years, this [world models, not LLMs] will be the dominant model for AI architectures, and nobody in their right mind would use LLMs of the type that we have today,” the 65-year-old said last month at a symposium at the Massachusetts Institute of Technology. 

LeCun has been talking to associates about creating a startup focused on world models, recruiting colleagues and speaking to investors, The Wall Street Journal previously reported. A world model learns about the world around it by taking in visual information, much like a baby animal or young child does, versus LLMs, which are predictive models based on vast databases of text."

 

....

[WayWardCloud]

I'd like to hear what exactly he's been right about for 40 years. Looks to me like he led Facebook nowhere and now Mark is left trying to catch up with everyone else even though he had the foresight to start an AI division at his company a whole decade ago.

Edited November 17, 2025 by WayWardCloud

[Libs]

34 minutes ago, WayWardCloud said:

I'd like to hear what exactly he's been right about for 40 years. Looks to me like he led Facebook nowhere and now Mark is left trying to catch up with everyone else even though he had the foresight to start an AI division at his company a whole decade ago.

 

He's right about LLM models being a dead-end. This is what Gary Marcus has been saying, loudly, and he's gradually being vindicated. 

[John Hjorth]

The Economist - Leaders [November 13^th 2025] : How markets could topple the global economy - If the AI buble bursts, an unusual recession could follow.

 

- - - o 0 o - - -

 

Interesting read, I would say. Food for thought, at least to me, because of owning BN. Where such deals just get bigger and bigger by now related to it one way or another, but likely primarily involving client money.

 

- - - o 0 o - - -

 

Article attached.

The Economist - How markets could topple the global economy - 20251113 - 20251117.pdf

[brobro777]

Boy there sure seems to be a lot of funboys flapping their gums about AI bubble/top

 

i never heard so much jibba about tops in 1999/2000 or 2006/2007

 

Maybe theres a chance for a squeeze the shorts rally into a very merry Christmas 

 

 

 

[rogermunibond]

@beerbaron have you read anything on the current attrition rate of Nvidia H200 (Hopper) GPUs?  when you buy a 1000 what percentage fail per year?  Anyone know if this has improved, stayed the same or worsened?

[Spekulatius]

10 hours ago, brobro777 said:

Boy there sure seems to be a lot of funboys flapping their gums about AI bubble/top

 

i never heard so much jibba about tops in 1999/2000 or 2006/2007

 

Maybe theres a chance for a squeeze the shorts rally into a very merry Christmas 

 

 

 

I guess you were not around in 1999/2000 then because there sure was a lot of talk about the bubble back then.

2006/2007 was different because is wasn’t a stock market valuation bubble , it was the financial system undermined  by a real estate bubble popping and unsound practices.

[Dalal.Holdings]

We all know that capital doesn't get misallocated on euphoric projects like these, especially when the Federal gov't is also slamming on the accelerator ...

[rogermunibond]

[Spekulatius]

Another interesting piece of information (if true)- upstarts overwhelmingly use Chinese AI models for two reason:

1) leaner and meaner ( cheaper to run)

2) They are open source rather than closed 

 

Sounds to me like OpenAI and the like have an issue as it’s likely they can never charge enough to pay for their Opex and Capex spent.

 

Same happened with telecom equipment and components which went to China in 2001-2003 because they were good enough and much cheaper to produce there. 

 

Edited November 17, 2025 by Spekulatius

[beerbaron]

10 hours ago, rogermunibond said:

@beerbaron have you read anything on the current attrition rate of Nvidia H200 (Hopper) GPUs?  when you buy a 1000 what percentage fail per year?  Anyone know if this has improved, stayed the same or worsened?

NVidia does does disclose MTBF but it's mentioned that they have a median life around 6 years. The problem with GPU is that it's all parallel work. So if you lose  a piece, the whole round of compute might be lost. In a way it's kind of the same thing as wireless networks (packet collision when two client emit at the same time), below a certain level and it is quite manageable but above another limit and the whole thing becomes useless. You can obviously mitigate that by mixing compute between GPU or having some redundancy which is what everybody is doing but then, you add a load on the telecom side of things.

In fact as I come to think of it, I'd rather invest in some next gen telecom for GPU cluster than the GPU themselves. Demand should scale as much but with a much better reliability and still has room to improve for quite a while.

 

BeerBaron

 

 

Edited November 18, 2025 by beerbaron

[brobro777]

10 hours ago, Spekulatius said:

I guess you were not around in 1999/2000 then because there sure was a lot of talk about the bubble back then.

2006/2007 was different because is wasn’t a stock market valuation bubble , it was the financial system undermined  by a real estate bubble popping and unsound practices.

Yea bro I was around back then and what I remember is Nasdaq cracking 50% from 5000 in march to 2500 in dec 2000 and people still not openly declaring it a bubble, not like now where prominent guys like Bill Gates, Altman, and the economist are declaring AI bubble now (https://www.economist.com/leaders/2025/11/13/how-markets-could-topple-the-global-economy). I remember there were guys saying it would all bounce back, fueled by stuff like two surprise interest rates from Greenspan in Jan 2001 and April 2001. My memory is that only after Nasdaq crashed 70% from 5000 in spring 2000 to less than 1500 in September 2001 that big prominent guys finally declared dot com bubble

 

as for the 2007-2009 bust, you could say that was a bubble in reckless mortgages...

[investmd]

Thesis: AI can be a catalyst for Large Pharma. 

 

AI for drug discovery has long been an obvious application for AI in healthcare. However, instead of early stage biotech being the big winners, it might be the boring old steady Big Pharma like MRK and SNY that are currently trading at P/E ratios of <15 and have steady 4-5% dividend yields. 

 

AI has the potential to drastically decrease cost of both drug discovery & development (clinical trial cost) thus driving down spend at Big Pharma. Then these massive, stable conglomerates can see a significant multiples expansion.  Seems to be upside with downside protection? 

Edited November 18, 2025 by investmd

[gfp]

2 hours ago, investmd said:

Thesis: AI can be a catalyst for Large Pharma. 

 

AI for drug discovery has long been an obvious application for AI in healthcare. However, instead of early stage biotech being the big winners, it might be the boring old steady Big Pharma like MRK and SFY that are currently trading at P/E ratios of <15 and have steady 4-5% dividend yields. 

 

AI has the potential to drastically decrease cost of both drug discovery & development (clinical trial cost) thus driving down spend at Big Pharma. Then these massive, stable conglomerates can see a significant multiples expansion.  Seems to be upside with downside protection? 

 

I don't know who the winners will be (all of them?) but Alphabet's Isomorphic labs and a startup called Insilico Medicine seem to be pretty well positioned.  Just reading about DeepMind's AlphaFold over the weekend was pretty cool.

 

Isomorphic Labs (Alphabet) has current partnerships with Lilly and Novartis.  Lilly isn't one of the single digit PE firms of course but a great company.

 

Insilico Médecine has partnerships with quite a few big pharma firms like Lilly, Sanofi, Pfizer, etc..

 

It's going to be an exciting time in that field but it's beyond my pay grade to figure out winners.  But I like that its good for Google.