Artificial Intelligence

[rogermunibond]

https://www.cryptopolitan.com/investors-flock-to-cds-as-ai-raises-concerns/

 

CDS trading on highly rated tech firms is up.  Creditors want protection, seems rational for now.

[Dalal.Holdings]

https://www.cnbc.com/2025/12/17/oracle-stock-blue-owl-michigan-data-center.html

 

Uh oh...

[NnnnotSoSmart]

Micron just very reported strong earnings.  

They claim on slide 27 that AI is benefitting the company. No dollar amounts mentioned.

Suspect we'll see more of these types of slides from corporations in the future.

Suppose the proof is in the pudding...will it ultimately benefit the bottom line?  Time will tell.

 

 

Edited December 17, 2025 by NnnnotSoSmart

[Dalal.Holdings]

Quote

I will be pushing for a moratorium on the construction of data centers that are powering the unregulated sprint to develop & deploy AI. The moratorium will give democracy a chance to catch up, and ensure that the benefits of technology work for all of us, not just the 1%.

 

Leftists in America and Europe make Xi very very happy

[sholland]

On 12/5/2025 at 7:47 PM, Spekulatius said:

The brute force approach is almost always beaten by smarter and more efficient approaches. We are very far away from superintelligence and the current LLM’s don’t reason at all.
I do think weaponization of AI and robots is probably close ( a couple of years way) and China could well have the lead there, due to mass manufacturing and good enough AI.  You don’t need your terminators to be all that smart either.

I think @Spekulatius is right and I am wrong.  The human brain consumes ~20W of power.  An Nvdia H100 consumes about ~1000W for similar performance.  Seems highly likely that chips will be more efficient than the human brain someday.  Thanks for the pushback.

[rogermunibond]

I think you need to compare the human brain to a finished LLM say the Gemini 3.0 Flash or ChatGPT o3 mini.  The smaller LLMs that can function without thousands of H100s for inference.  Some of these reduced token LLMs can even work very quickly on a smartphone (locally).  That's the comparison.

 

When you train an LLM over say 6 months on 1000 H100s compare that to training a human brain over 22 years and the caloric/power (3000 calories per day) requirement to get that individual educated.  4,380 hours X 1000 H100 X 1,000 kwh vs 190,720 hours X 522 kwh.

 

4,380,000,000 watts vs 99,555,480 watts

 

4.38 Gwh vs about 100 Mwh

 

So I guess the human is still more power efficient but certainly takes  a lot longer.

[djokovic1]

I am interested if others here find AI useful for investment research. For my own research process, I have found AI mostly useless. I just go to source material like fillings and annual reports to start if something interests me. It seems contrary to the popular narrative that AI is changing the world so I am curious if I am just being a dinosaur. If others find AI useful in the research process, I would love to hear how you use it? I should add that I mostly look at small cap names, so maybe that adds to it.

[bargainman]

On 12/21/2025 at 3:47 PM, sholland said:

I think @Spekulatius is right and I am wrong.  The human brain consumes ~20W of power.  An Nvdia H100 consumes about ~1000W for similar performance.  Seems highly likely that chips will be more efficient than the human brain someday.  Thanks for the pushback.

Funny that.  I just read this which is tangentially but still related.

 

The Bitter Lesson

 

 

[bargainman]

On 12/22/2025 at 5:58 AM, rogermunibond said:

I think you need to compare the human brain to a finished LLM say the Gemini 3.0 Flash or ChatGPT o3 mini.  The smaller LLMs that can function without thousands of H100s for inference.  Some of these reduced token LLMs can even work very quickly on a smartphone (locally).  That's the comparison.

 

When you train an LLM over say 6 months on 1000 H100s compare that to training a human brain over 22 years and the caloric/power (3000 calories per day) requirement to get that individual educated.  4,380 hours X 1000 H100 X 1,000 kwh vs 190,720 hours X 522 kwh.

 

4,380,000,000 watts vs 99,555,480 watts

 

4.38 Gwh vs about 100 Mwh

 

So I guess the human is still more power efficient but certainly takes  a lot longer.

 

Yeah, but the human brain isn't improving exponentially and doesn't have a boatload of researchers and capital incentivized to make it better, faster, cheaper and more more more.  Don't look now, look next year, year after, and after ...

[Spekulatius]

On 12/22/2025 at 2:58 PM, rogermunibond said:

I think you need to compare the human brain to a finished LLM say the Gemini 3.0 Flash or ChatGPT o3 mini.  The smaller LLMs that can function without thousands of H100s for inference.  Some of these reduced token LLMs can even work very quickly on a smartphone (locally).  That's the comparison.

 

When you train an LLM over say 6 months on 1000 H100s compare that to training a human brain over 22 years and the caloric/power (3000 calories per day) requirement to get that individual educated.  4,380 hours X 1000 H100 X 1,000 kwh vs 190,720 hours X 522 kwh.

 

4,380,000,000 watts vs 99,555,480 watts

 

4.38 Gwh vs about 100 Mwh

 

So I guess the human is still more power efficient but certainly takes  a lot longer.

I think a lot of the interference will be on mobile devices eventually (becoming your personal AI Assistent ) and that will force power efficiency. Only complex task (by future standards) will run on the datacenter cloud.

I think this all will happen because efficiency gains are likely. The first computer running with tubes also consumed huge amounts of powers just a do simple things like multiplications. AI is going to become a race of efficiency where not the fastest and best solution is going to win, but whoever brings a useful solution for day to day problems into the wind of people at low cost.

 

There may be a high end market for some applications but the bleeding edge won’t be where the biggest part of the TAM is.

Edited December 24, 2025 by Spekulatius

[bargainman]

On 12/24/2025 at 12:14 PM, Spekulatius said:

I think a lot of the interference will be on mobile devices eventually (becoming your personal AI Assistent ) and that will force power efficiency. Only complex task (by future standards) will run on the datacenter cloud.

I think this all will happen because efficiency gains are likely. The first computer running with tubes also consumed huge amounts of powers just a do simple things like multiplications. AI is going to become a race of efficiency where not the fastest and best solution is going to win, but whoever brings a useful solution for day to day problems into the wind of people at low cost.

 

There may be a high end market for some applications but the bleeding edge won’t be where the biggest part of the TAM is.

 

I assumed that you meant to say inference not interference.  it will be interesting to see how things pans out.  Local models are definitely still not powerful enough to handle a lot of the tasks or at least not reliably.  also a lot of the infrastructure is built on python, and that is something that's going to take a long time to get ported over to native local systems, if it does it all.

 

There are a lot of races going on at the moment.  Getting models to run locally is generally a very heavy lift vs just calling a huge model up on a server running on some massive GPUs farm.  But these things do move in cycles.

[UK]

https://archive.is/c8H1Q

 

An AI revolution in drugmaking is under way

It will transform how medicines are made—and the industry itself

[Longnose]

On 12/23/2025 at 11:10 PM, djokovic1 said:

I am interested if others here find AI useful for investment research. For my own research process, I have found AI mostly useless. I just go to source material like fillings and annual reports to start if something interests me. It seems contrary to the popular narrative that AI is changing the world so I am curious if I am just being a dinosaur. If others find AI useful in the research process, I would love to hear how you use it? I should add that I mostly look at small cap names, so maybe that adds to it.

I use it all the time. (you just always have to fact check anything when it comes to numbers)

 

One of my most preferred ways to use it is more on the thesis standpoint. I use TIKR for screens and numbers mostly.  For me I find a lot of value in opinions. Ive subscribed to SA for years just to have access to the comments on articles. I dont care what most articles say but i want to know what other educated (and uneducated) people are agreeing with or arguing against.  Recently, ive been using AI to take that opinion side of things. If im bullish i want to argue my case against someone. I find a lot of value in someone or something playing devils advocate with me and AI is really good at that. So i usuually know my own numbers and dont let the AI quote numbers at me. But more in debating thesis.  When im exploring industries im less familiar with ill have gemini do deep reseach on the industry and give me good comprehensive reports that can be quite educational. 

 

I view it like im the CIO and I have an Sr Analyst to delegate things i want to research to. Make me a 20 page report on the history of company "X"

or

"go read the last 15 earnings call transcripts look for what management has said they are going to do and then compare that with what they actually succeeded on. After your written summary put the execution list in a table with completed not completed and transcript action was stated in and if completed when reported completed. "

 

I dont have a problem reading and because AI makes shit up sometimes you do need to fact check. But its getting better and better to the point that 90+ % of the time its right so yea its helpful 

 

At least for me. 

[gfp]

12 hours ago, Longnose said:

I use it all the time. (you just always have to fact check anything when it comes to numbers)

 

One of my most preferred ways to use it is more on the thesis standpoint. I use TIKR for screens and numbers mostly.  For me I find a lot of value in opinions. Ive subscribed to SA for years just to have access to the comments on articles. I dont care what most articles say but i want to know what other educated (and uneducated) people are agreeing with or arguing against.  Recently, ive been using AI to take that opinion side of things. If im bullish i want to argue my case against someone. I find a lot of value in someone or something playing devils advocate with me and AI is really good at that. So i usuually know my own numbers and dont let the AI quote numbers at me. But more in debating thesis.  When im exploring industries im less familiar with ill have gemini do deep reseach on the industry and give me good comprehensive reports that can be quite educational. 

 

I view it like im the CIO and I have an Sr Analyst to delegate things i want to research to. Make me a 20 page report on the history of company "X"

or

"go read the last 15 earnings call transcripts look for what management has said they are going to do and then compare that with what they actually succeeded on. After your written summary put the execution list in a table with completed not completed and transcript action was stated in and if completed when reported completed. "

 

I dont have a problem reading and because AI makes shit up sometimes you do need to fact check. But its getting better and better to the point that 90+ % of the time its right so yea its helpful 

 

At least for me. 

 

Yeah I agree with all of this and I am using it more and more every day.  I just subscribed to Gemini Pro (it's on sale right now for like $100 for a year and you can invite several people to share the account if your family is into it. Mine isn't).

 

You can use @YouTube extension to summarize a podcast video, focus on specific questions, find stuff it would take you a while to find instantly.  You can upload around 1500 pages of files and ask it anything you want.  Devils advocate / getting into both sides of an argument really thoroughly.  It even had full access to everything said at Bay County (FL) board of commissioners meetings because it has access to all the minutes of every meeting.  Just crazy how quickly these AI agents are progressing to a full on jr. analyst where it was basically a LLM parlor trick not that long ago.

 

I'm still learning about "Gems" which are basically little macros you can write to do the same thing to different sources over time.

 

I'm not as in to voice mode but a lot of people talk to Gemini or Grok and debate things out verbally when they are driving or taking a walk.  I'm not there yet.  A lot of Tesla drivers are talking to Grok instead of listening to music I guess.

[MungerWunger]

The AI revolution is here. Will the economy survive the transition?

Michael Burry, Dwarkesh Patel, Patrick McKenzie, and Jack Clark

https://post.substack.com/p/the-ai-revolution-is-here-will-the?utm_campaign=post-expanded-share&utm_medium=post viewer&triedRedirect=true

[djokovic1]

On 1/9/2026 at 4:15 PM, Longnose said:

"go read the last 15 earnings call transcripts look for what management has said they are going to do and then compare that with what they actually succeeded on. After your written summary put the execution list in a table with completed not completed and transcript action was stated in and if completed when reported completed. "

Thanks for your thoughts and how you use it. While I don't disagree, a counter point.

 

My real conviction on Fairfax has come from reading 25 years of its annual reports myself like a story starting from 2000 (not by having AI summarise it). In addition reading @Viking's amazing 800 page compendium. 

 

I don't think I would have the same conviction if I didn't do the hard work. Just like reading an AI summary of a book is nothing compared to reading the actual book. Your brain picks up on a lot of things directly and indirectly going through that process. 

 

This is not to say AI can't add value as you suggest by using it smartly especially in the initial assessment phase to see if a company is worth a deep dive. 

[UK]

12 hours ago, MungerWunger said:

The AI revolution is here. Will the economy survive the transition?

Michael Burry, Dwarkesh Patel, Patrick McKenzie, and Jack Clark

https://post.substack.com/p/the-ai-revolution-is-here-will-the?utm_campaign=post-expanded-share&utm_medium=post viewer&triedRedirect=true

Thanks for sharing!

[formthirteen]

On 1/9/2026 at 6:15 PM, Longnose said:

"go read the last 15 earnings call transcripts look for what management has said they are going to do and then compare that with what they actually succeeded on. After your written summary put the execution list in a table with completed not completed and transcript action was stated in and if completed when reported completed. "

 

I dont have a problem reading and because AI makes shit up sometimes you do need to fact check. But its getting better and better to the point that 90+ % of the time its right so yea its helpful 

 

Thanks for sharing. I tried your approach on CROX earnings call transcripts and was impressed by the detailed report I got and research.

[Longnose]

14 hours ago, djokovic1 said:

Thanks for your thoughts and how you use it. While I don't disagree, a counter point.

 

My real conviction on Fairfax has come from reading 25 years of its annual reports myself like a story starting from 2000 (not by having AI summarise it). In addition reading @Viking's amazing 800 page compendium. 

 

I don't think I would have the same conviction if I didn't do the hard work. Just like reading an AI summary of a book is nothing compared to reading the actual book. Your brain picks up on a lot of things directly and indirectly going through that process. 

 

This is not to say AI can't add value as you suggest by using it smartly especially in the initial assessment phase to see if a company is worth a deep dive. 

Totally get it. AND i agree with you. 

 

I view it as a good prescreener. does the AI summary warrant a deep dive worth going 800 pages deep? Or did the assistant say this company is full of shit? 

 

[djokovic1]

It was a good AI summary. The final summary from the free version:

 

Bottom Line

This is absolutely a public investment worth studying further, especially if you are drawn to:

• Insurance economics

• Capital allocation stories

• Underfollowed, complex businesses

It is not a “quick yes/no” stock — but the quality of the work alone justifies deeper diligence.

[djokovic1]

And I really liked AI's answer of reason's why not to do a deep dive on Fairfax (although I would recommend it )

 

Below are legitimate, non-trivial reasons why not doing a deep dive on Fairfax could be rational, even if the bull case is coherent and well-argued.

I’ll group them by time efficiency, investor fit, and structural issues.

---

1. It Is a Time-Intensive, High-Cognitive-Load Company

Reason to pass: Your research time is scarce, and Fairfax is unusually demanding.

Fairfax requires understanding:

• P/C insurance economics

• Float dynamics

• Fixed income duration management

• Capital allocation history

• Multiple geographies and private assets

• Accounting noise vs economic reality

This is not a company you can understand well in:

• 2 hours

• One earnings call

• A single valuation model

If your process favors:

• Simple business models

• Rapid screening and turnover

• Pattern recognition over deep synthesis

 Fairfax is a poor use of time, even if it’s a good business.

---

2. The Payoff to Insight May Be Limited From Here

Reason to pass: The mispricing may have narrowed enough that deep work no longer earns excess returns.

The document itself acknowledges:

• Stock is up ~3–4× since 2020

• Narrative has shifted from “turnaround” to “quality”

• Sell-side and ratings agencies are now warming

That suggests:

• The big informational edge may already be gone

• Future returns may track fundamentals more closely

If you require:

• Large asymmetric upside from discovery

• Situations where few investors are paying attention

Fairfax may now be more about execution and compounding, not misunderstanding arbitrage.

---

3. Conglomerates Often Stay Discounted — Even When They Improve

Reason to pass: The market may never reward Fairfax with the multiple implied by the bull case.

Structural headwinds:

• Complex reporting

• Mix of public + private + foreign assets

• Hard to benchmark peers cleanly

• Earnings volatility from mark-to-market accounting

History shows:

• Conglomerate discounts can persist for decades

• Even excellent capital allocators don’t always get rerated

If your thesis requires:

• Multiple expansion

• “Recognition” by the market

That recognition may not come, or may come very slowly.

---

4. Insurance Is Cyclical — and Timing Matters

Reason to pass: You may be late in the underwriting cycle.

Key risk:

• Current profitability is partly driven by a hard insurance market

• History shows underwriting margins mean-revert

Even if Fairfax is well run:

• Premium pricing can soften

• Cat losses can spike

• ROE can compress for years

If your style avoids:

• Cyclical businesses

• Situations where timing matters materially

Insurance companies can be psychologically and financially uncomfortable holds.

---

5. You Are Implicitly Betting on One Person

Reason to pass: Key-man risk is real.

Fairfax is deeply associated with:

• Prem Watsa

• His capital allocation judgment

• His cultural influence

Even with succession planning:

• Style shifts matter

• Capital allocation errors compound

• The moat here is human, not structural

If your process prefers:

• Systemic moats (network effects, switching costs)

• Businesses that thrive independent of leadership

Fairfax may violate your risk preferences.

---

6. The Best Case Requires Comfort With Volatility and “Ugly” Periods

Reason to pass: This will not behave like a smooth compounder.

Even in bull scenarios:

• Quarterly earnings can swing wildly

• Book value can dip due to markets

• Headlines may look bad even when economics are fine

If you:

• Mark to market emotionally

• Need narrative consistency

• Prefer steady optics

Fairfax can test discipline, even for long-term investors.

---

7. Opportunity Cost: There May Be Easier Wins

Reason to pass: Your effort-to-return ratio may be inferior elsewhere.

You could instead:

• Study a simpler insurer with cleaner disclosure

• Buy a higher-quality compounder with obvious moat

• Allocate to an index and spend time elsewhere

Deep dives only make sense when:

“The expected incremental return from understanding this deeply exceeds that of alternatives.”

That may or may not be true for you.

---

A Clean Decision Framework (Use This)

You should not deep dive Fairfax if 2 or more of these are true:

• ☐ I dislike complex, multi-moving-part businesses

• ☐ I need clear, recurring revenue visibility

• ☐ I rely on multiple expansion for returns

• ☐ I avoid cyclical industries

• ☐ I don’t want key-man exposure

• ☐ I don’t enjoy insurance accounting

• ☐ I already have enough ideas with higher conviction

If most boxes are checked, passing is not laziness — it’s discipline.

---

Bottom Line

Fairfax may be a good investment and still a bad use of your time.
Those are not contradictory statements.

Edited January 12 by djokovic1

[Libs]

On 1/5/2026 at 11:53 PM, UK said:

https://archive.is/c8H1Q

 

An AI revolution in drugmaking is under way

It will transform how medicines are made—and the industry itself

 

This deserves more attention. It's revolutionizing drug research. A friend of mine, a very senior guy in the industry, confirmed this  to me recently.

 

 

 

Edited January 13 by Libs

[Fly]

6 minutes ago, Libs said:

 

This deserves more attention. It's revolutionizing drug research. A friend of mine, a very senior guy in the industry, said confirmed this  to me recently.

 

 

 

 

What is the best way to invest in this? RXRX? Insilico's HK listing?

[formthirteen]

On 1/6/2026 at 9:53 AM, UK said:

An AI revolution in drugmaking is under way

It will transform how medicines are made—and the industry itself

 

Thank you. I wasn't aware of this development.

 

6 hours ago, Fly said:

What is the best way to invest in this? RXRX? Insilico's HK listing?

 

I would also be interested in hearing about investment ideas, for example, Dassault Systemes has MEDIDATA:

https://www.3ds.com/products/medidata

 

Quote

Discover the Industry Standard

Drawing on a quarter century of global corporation with customers, patients, and partners, MEDIDATA delivers unmatched results to power smarter treatments and healtier people.

    36,000+

    clinical trials
    11+ million

    patients
    70+ billion

    data points annually
    2,300+

    customers

MEDIDATA's AI-powered Patient, Data, and Study Experiences improve patient engagement through digital technologies, enhance data quality with actionable insights, and optimize clinical trial design with predictive analytics.

 

https://investor.3ds.com/static-files/0b6ce277-804d-4df7-868d-e5acfc9adb29

Also:

https://investor.3ds.com/news-releases/news-release-details/dassault-systemes-outperforming-third-quarter-confirming-full/#:~:text=Life Sciences software revenue increased,baseline due to Covid trials.

Edited January 13 by formthirteen

[UK]

2 hours ago, formthirteen said:

I would also be interested in hearing about investment ideas

 

Somebody earlier proposed to invest into big pharma because of this, but I am not sure it will not be this "Buffett elevator" thing longer time for them, of course positive in this case, but perhaps to be competed out longer term?

 

Edited January 13 by UK