Artificial Intelligence

[Spekulatius]

At this point, people just make shyte up.

 

One thing is for sure, AI is the biggest capital cycle of the last 50 years. If it’s really life changing remains TBD.

[MungerWunger]

 

Meta AI director failed to control her AI assistant 

Edited February 24 by MungerWunger

[Valuebo]

4 hours ago, Spekulatius said:

At this point, people just make shyte up.

 

One thing is for sure, AI is the biggest capital cycle of the last 50 years. If it’s really life changing remains TBD.

 

Yes, it's wild. It's a self-reinforcing echo chamber where people seemingly completely unfamiliar with the inner workings of large, complex, slow-moving businesses spin fairy tales about AI taking over. 

 

I reaaaally want to see the first major company overhauling half their internal operations in 2027 or 2028 because they decided to replace something like JIRA (or any other mission-critical infrastructure) in favor of some in-house made, half-assed, non-audited slop for which they have no proper documentation, no external support, no legal accountability and unreliable maintenance costs. Let alone financial players simply replacing COBOL for their mainframes. Not to mention the shitload of people you have to re-train, potential compatibility issues, management career risk, impact on operations and sales, ... It's simply laughable and that would still be true even if you could perfectly code a decent replacement of existing systems and implement them with minimal risk. It's just not going to happen overnight.

 

Have any of these people who believe SaaS is already going to lose a decent chunk of their revenue in the next few years ever actually been inside the departments of these companies working with this software? Do they realize how hard it is even today to get Sally from sales to properly fill in the expense reports? 

 

Even if AI suddenly became somewhat capable of deterministic outcomes without real risk of absolute catastrophic faults or hallucinations, you are not going to see any major moves towards these tools for most of the big companies in the next 2-3 years. For this to even happen over the longer term, the actual claims of AI companies still have to come to fruition as well as it currently is still all based on open source knowledge and statistical guesswork with a lot of human intervention to make barely working and half assed projects. There is general expectancy that more data and training of models will keep the evolution (and pace) of AI intact, but the jury is still out. So far, it largely remains make-believe if you ask me. Good tools exist already, but they make up a small part of what makes most successful SaaS companies work.

 

Sure, margins can take a hit as the market rebalances but I don't believe we should expect earthshattering movements for all companies here either. In fact, I believe the current crisis will finally be the push many of these SaaS 1.0 companies need to right the ship in terms of costs bloat, SBC, ... 

Edited February 24 by Valuebo

[UK]

'Some people get rich studying artificial intelligence. Me, I make money studying natural stupidity'. Carl Icahn.

[UK]

https://archive.is/BaIXK

 

After a three-year love affair with anything related to artificial intelligence, U.S. investors are flocking to the factory owners, fast-food restaurants and commodity companies that have seemingly strong odds of surviving the technological revolution intact.

 

Call it the AI immunity trade, HALO—for “heavy assets, low obsolescence”—or just another iteration of the jitters that have periodically rippled through markets since the AI investing boom began. The winners include McDonald’s, Exxon Mobil XOM -2.44%decrease; red down pointing triangle and tractor maker Deere DE 0.07%increase; green up pointing triangle. Left behind are the perceived potential victims of the AI revolution, a list that has ranged from wealth managers to software firms.

[UK]

https://archive.is/Osofk

 

The AI productivity boom is not here (yet)
Artificial intelligence is improving fast. Its effect on output, not so much

 

All this signals a deeper flaw in the argument that AI is powering a productivity boom. Such improvements are usually made not just when workers use a new tool more often, but when firms reorganise production around it. Early factories became only a little more efficient when steam engines were replaced with electric motors; the real revolution came decades later after floor plans were redesigned to make the most of electric power. More recently, productivity growth was a disappointment for years after personal computers became widespread. It accelerated only once firms implemented business models that exploited the technology to its full potential. Much of America’s productivity revival in the 1990s came not from Silicon Valley itself but from retail, where computers transformed logistics and inventory management.

 

[UK]

 

[frommi]

When you like employees that make shit up >50% of the time, go for agentic AI . 
 

Recent benchmarks (like OpenAI's SimpleQA) specifically target "real-world" facts that are difficult or obscure. When faced with these:
Top-Tier Models (GPT-4o, Claude 3.5): Correct only about 38% to 40% of the time. The rest of the responses are either "I don't know" or confident hallucinations.
Reasoning Models (o1, o3): Perform better (roughly 60%+ accuracy) because they "double-check" their internal logic, but they still hallucinate on roughly 1 in 3 difficult factual questions.

 

And this is embedded in the way the models are created, that doesnt get better from model to model.

Edited February 24 by frommi

[Spekulatius]

This could best interesting :

https://apnews.com/article/anthropic-hegseth-ai-pentagon-military-3d86c9296fe953ec0591fcde6a613aba

 

Restriction of use are pretty common. Had our legal guy sign some for some hardware from Japan that was part of a piece of equipment we purchase.

[rogermunibond]

https://www.cnbc.com/2026/02/25/trump-tech-ai-data-center-electricity-price-pledge.html

 

Agreement coming soon for hyperscalers to build their own electrical generation for data centers.

[gfp]

1 minute ago, rogermunibond said:

https://www.cnbc.com/2026/02/25/trump-tech-ai-data-center-electricity-price-pledge.html

 

Agreement coming soon for hyperscalers to build their own electrical generation for data centers.

 

Gotta be a positive for Bloom since nobody else has availability.  Don't suppose the president thought through to natural gas prices since he's a pretty first order kinda guy.  Luckily there is plenty of gas in the US

[nsx5200]

 

Ran across this in my feed.  This is a PhD/researcher's POV on the practical and theoretical limitations of current AI(LLM) systems.  It talks about different cases where the models can break and degrade, especially if trained from output of LLM output.

 

This has deep implication since a lot of these AI system's output is put on the internet, and the same content on the internet is used to train future AI systems.  In addition, it goes into difference between syntax and semantics.  The researcher claims, with evidence, that current AI LLM operates on syntax(words/symbols) without true understanding(semantics) and so there will be fundamental flaws in how these LLM tries to solve certain type of problems.  For me, this was a bit of thought-provoking, as IMHO, many people, like these LLMs, walk through life operating mostly on syntax as well.  Many of these syntax-based jobs will be replaced by these LLMs.

 

Probably a bit more academic than most of those content produced by people with financial background, but IMHO, has more solid evidence, albeit a bit more theoretical, than the "I vibe coded, and it produced something somewhat working" evidence.

[Libs]

13 hours ago, nsx5200 said:

 

Ran across this in my feed.  This is a PhD/researcher's POV on the practical and theoretical limitations of current AI(LLM) systems.  It talks about different cases where the models can break and degrade, especially if trained from output of LLM output.

 

This has deep implication since a lot of these AI system's output is put on the internet, and the same content on the internet is used to train future AI systems.  In addition, it goes into difference between syntax and semantics.  The researcher claims, with evidence, that current AI LLM operates on syntax(words/symbols) without true understanding(semantics) and so there will be fundamental flaws in how these LLM tries to solve certain type of problems.  For me, this was a bit of thought-provoking, as IMHO, many people, like these LLMs, walk through life operating mostly on syntax as well.  Many of these syntax-based jobs will be replaced by these LLMs.

 

Probably a bit more academic than most of those content produced by people with financial background, but IMHO, has more solid evidence, albeit a bit more theoretical, than the "I vibe coded, and it produced something somewhat working" evidence.

Thank you. Very good video. Gary Marcus, a lead skeptic on LLM scaling, would agree with this.

 

I'm struggling as an investor to come to grips with AI. On the one hand, having lived through multiple hype cycles, and having avoided / profited by their ultimate collapses, I'm now hard-wired to be  skeptical of this one. I'm quite sure,for example, that scaling LLM's can not lead to reasoning (AGI), as this video points out. But.....

 

.....but people I respect* are testifying to the power of current LLM's - in their personal and business lives. So maybe AGI is not necessary for true disruption ( as I had assumed).

 

I'm settling on this, for now: 

 

1) AGI is not necessary for disruption. (if we ever get to true AGI it will be massively disruptive).

2) LLM's will still have to get a lot better to be as disruptive as AI maximalists believe.

 

So I'm looking for evidence of LLM models plateauing (or not).

 

Interested in other's thoughts for sure.

 

*GFP for example- once proud value guy now following Jordi Visser, advocating for AI disruption, and using charts!

 

 

 

[Milu]

1 hour ago, Libs said:

Thank you. Very good video. Gary Marcus, a lead skeptic on LLM scaling, would agree with this.

 

I'm struggling as an investor to come to grips with AI. On the one hand, having lived through multiple hype cycles, and having avoided / profited by their ultimate collapses, I'm now hard-wired to be  skeptical of this one. I'm quite sure,for example, that scaling LLM's can not lead to reasoning (AGI), as this video points out. But.....

 

.....but people I respect* are testifying to the power of current LLM's - in their personal and business lives. So maybe AGI is not necessary for true disruption ( as I had assumed).

 

I'm settling on this, for now: 

 

1) AGI is not necessary for disruption. (if we ever get to true AGI it will be massively disruptive).

2) LLM's will still have to get a lot better to be as disruptive as AI maximalists believe.

 

So I'm looking for evidence of LLM models plateauing (or not).

 

Interested in other's thoughts for sure.

 

*GFP for example- once proud value guy now following Jordi Visser, advocating for AI disruption, and using charts!

 

 

 

Ya I'm not too sure what to think either. I'm currently in a mostly wait and see approach. There is a lot of seemingly good value out there at the moment in SAAS stocks but I think the future is just too unpredictable for me to take much action. I know this is always the case and I'm usually quite decisive in large stock drawdowns like we see in certain enterprise software names. This time feels different, famous last words!

 

I was mostly in the camp of AI is going to change everything, this had me feeling that all tech firms (outside of the hyperscalers) are doomed. I've now dialled that back a little bit but my current investing framework is in a somewhat barbell manner of just owning the hyperscalers on one end as I think they will benefit massively from an AGI world, and at the other end owing companies like Dominos, Chipotle, Ferrari who should be impacted less. It's the mid level tech companies that I feel could struggle. 

 

 

I use claude a lot each day for my job and it really has at least doubled both the speed and quality of my work, so I can see the hype is real. Even if things were frozen today the current LLM's are transformational for some jobs (coders, consultants). So it's harder for me to see it as misplaced hype when my own two eyes confirm things each day. How this applies to the wider world of other industries, who knows.

Edited February 26 by Milu

[frommi]

Funnily i think its the opposite the real ones doomed are the hyperscalers and the biggest beneficiaries are the capex light software businesses that profit from the infrastructure laid and collect a royalty with their AI enabled software. Every software vendor where i listened to the calls can switch the LLM's with a click and gets paid by usage. OpenAI is already struggling, burning money and losing paying customers. 10 years from now compute will be very cheap because of the massive over-building of capacity right now. 
Dumping all your FCF into low ROI investments will likely have a price, ie valuation compression. That will hit the whole index because all the big trillion dollar companies are in this game and they make up 30-40% of the indices. Deprecation will go up a lot next year for the hyperscalers and forward, surpressing earnings. 

[yesman182]

So todays story line is Block cutting their workforce by 40%. Since they are cutting workers it trades up 15%. Meanwhile all other payments and sas companies go down. At the same time Dorsey says that all companies will come to the realization that they have too much staff and will make a similar staffing adjustment at some point this year. interesting space to pay attention too 

[rogermunibond]

40 minutes ago, yesman182 said:

So todays story line is Block cutting their workforce by 40%. Since they are cutting workers it trades up 15%. Meanwhile all other payments and sas companies go down. At the same time Dorsey says that all companies will come to the realization that they have too much staff and will make a similar staffing adjustment at some point this year. interesting space to pay attention too 

 

Block is something of an outlier because they expanded so much and Dorsey probably didn't have his eye on the headcount.

[NnnnotSoSmart]

22 hours ago, frommi said:

Funnily i think its the opposite the real ones doomed are the hyperscalers and the biggest beneficiaries are the capex light software businesses that profit from the infrastructure laid and collect a royalty with their AI enabled software. Every software vendor where i listened to the calls can switch the LLM's with a click and gets paid by usage. OpenAI is already struggling, burning money and losing paying customers. 10 years from now compute will be very cheap because of the massive over-building of capacity right now. 
Dumping all your FCF into low ROI investments will likely have a price, ie valuation compression. That will hit the whole index because all the big trillion dollar companies are in this game and they make up 30-40% of the indices. Deprecation will go up a lot next year for the hyperscalers and forward, surpressing earnings. 

Chris Bloomstran's recent client letter dedicates 16 pages starting at page 60 of 200 to his skeptical AI thesis:

 

"Joni Mitchell penned Both Sides Now sixty years ago and only days after Warren Buffett
gained control of Berkshire Hathaway. Warren and his business partner of decades, Charlie Munger,
shared skepticism of AI, Charlie noting, “I think old-fashioned intelligence works pretty well.”

 

A section in this year’s letter (pages 60-76) attempts a skeptical case about AI, particularly of the illogic of expecting even a reasonable investment return on the billions, and what will soon be trillions, of dollars of capital being poured into the endeavor. We are witness to a growing reliance on debt, both on balance sheet and off, circular finance and accounting gimmickry. It looks like a throwback may be afoot to the telecommunications and fiber-optic cable boom and to earlier capital cycle bubbles. As with every transformative capital cycle, be it canals, railroads, electricity or the internet, while society may benefit in the end, early investors? Not so much. This one is apt to end badly."

 

https://static.fmgsuite.com/media/documents/2a3a9bb6-ddd0-4384-8175-99736bed4daf.pdf

 

Edited February 27 by NnnnotSoSmart

[Spekulatius]

On 2/26/2026 at 11:16 AM, Libs said:

Thank you. Very good video. Gary Marcus, a lead skeptic on LLM scaling, would agree with this.

 

I'm struggling as an investor to come to grips with AI. On the one hand, having lived through multiple hype cycles, and having avoided / profited by their ultimate collapses, I'm now hard-wired to be  skeptical of this one. I'm quite sure,for example, that scaling LLM's can not lead to reasoning (AGI), as this video points out. But.....

 

.....but people I respect* are testifying to the power of current LLM's - in their personal and business lives. So maybe AGI is not necessary for true disruption ( as I had assumed).

 

I'm settling on this, for now: 

 

1) AGI is not necessary for disruption. (if we ever get to true AGI it will be massively disruptive).

2) LLM's will still have to get a lot better to be as disruptive as AI maximalists believe.

 

So I'm looking for evidence of LLM models plateauing (or not).

 

Interested in other's thoughts for sure.

 

*GFP for example- once proud value guy now following Jordi Visser, advocating for AI disruption, and using charts!

 

I think some of the sceptic claim that LLM have a ceiling in term of how good they can be  and it may not be much higher than what we currently have at hand. The reason is that LLM’s are basically just guessing machines based on context once they ingested all the publicly available data, and squeeze juice out of the algos, they are pretty much done improving.

 

Adding 10x of the computing will probably not improve results much which means that all those folks that invested in this infrastructure are pretty much screwed.

 

Now, it may not be quite that bad, as you can have running the LLM trained on non public proprietary data (owned by corporation ) who create their own models and apps, based on practical use cases. For this, you don’t need extremely advanced models but it may take time to get the date cleaned enough to be able to train and then apply this to practical use cases in many domains. 
 

I

Edited February 28 by Spekulatius

[NnnnotSoSmart]

Howard Marks

 

AI Hurtles Ahead

 

https://www.oaktreecapital.com/insights/memo/ai-hurtles-ahead

 

 

 

[Spekulatius]

DOD is now  toxic to do business with as they deemed Anthropic to be “supply chain risk”. “Supply chain risk” means black listed like Chinese co. that means to federal agency or even contractor working for a federal agency can use them. It’s the nuclear option.

Now the problem is - who wants to do business with the DOD?

https://thehill.com/policy/defense/5759630-pentagon-designates-anthropic-risk/
 

[elliott]

"AI" software replacing "traditional" software does not make much sense to me. 

I see the disruption in the application layer becoming obsolete.

50 years ago finance, procurement, sales... these functions existed without relying on SAP or SalesForce.

The same kind of transformation can happen again.

[Spekulatius]

11 hours ago, elliott said:

"AI" software replacing "traditional" software does not make much sense to me. 

I see the disruption in the application layer becoming obsolete.

50 years ago finance, procurement, sales... these functions existed without relying on SAP or SalesForce.

The same kind of transformation can happen again.

How do you envision this to happen? LLM can’t do accounting, since they not deterministic, so the application layer will persist. LLM’s can assist users and there is a lot of value that can created by making data easier to extract.

Edited March 1 by Spekulatius

[elliott]

14 hours ago, Spekulatius said:

How do you envision this to happen? LLM can’t do accounting, since they not deterministic, so the application layer will persist. LLM’s can assist users and there is a lot of value that can created by making data easier to extract.

 

I didnt mean to say that LLMs will do the accounting or whatever else is needed.

I obviously dont know what will happen, but today applications are UI + business logic + data, and on these fronts I see the following developments.

 

Agents, or rather agentic workflows. These are applications based on LLMs. One pattern, here, is "tools". For example, an LLM doesnt know what time it is when you talk to it (just as it doesnt do accounting), but have a tool (e.g. an API) that tells which time it is, and whenever the LLM needs the time, it will say so in its answer, then an intermediate layer "calls" the time API and prompts back the LLM with your initial prompt + the time. Now, the LLM knows the time and can repply accordingly.

 

Data. Many companies are building large data repositories sourcing data from a variety of systems, external and internal, such as operational applications, maybe aligning the data to business processes, and then they make it available company-wide for any consumer. A "consumer" can be a model to improve yield in a factory for a given product, or even another operational application.

 

Now, could applications - as they are conceived today - be replaced by APIs running the business logic, using large common repositories of data, with agents and LLMs acting as a sort of UI for human users, and even as actors and orchestrators at the same time? 

I am already seeing this for some use-cases. To replace a whole applicaton such as SalesForce, however, is a whole different story. But who knows where technology will be in 10 or 15 years to say?

 

 

 

Edited March 1 by elliott

[Spekulatius]

Yes, apps already have API. Thats how they interact with other applications, at least the modern ones. I think any of the LLM agents will be build by the SaaS companies who build the current apps l it’s their market to lose.

 

Some of these AI maximists think they can just extract all the data over API’s and put the business logic in agents and the SaaS companies would just allow them to do that. First of all, business logic isn’t that simple and second the SaaS companies won’t give up their turf without a fight and taking a first shot at creating and selling the agents themselves.