Artificial Intelligence

[gfp]

this guy is a very good follow for a daily summary of everything important that happened on the frontiers of AI.  you may recognize him from podcasts like all-in and similar stuff.  maybe reading these daily summaries every day for a month will help some of you get into the present and move beyond "a lawyer once submitted a chatGPT generated brief in 2023 using ChatGPT-3.5 and it made shit up and that is a useful data point for me today."

 

https://x.com/alexwg/status/2021620145076998334

 

[frommi]

Yeah looks like we are all doomed because:
Facebook now lets users

animate profile pictures with AI

, while YouTube rolled out

text-to-playlist generation

.

[gfp]

7 minutes ago, frommi said:

Yeah looks like we are all doomed because:
Facebook now lets users

animate profile pictures with AI

, while YouTube rolled out

text-to-playlist generation

.

 

Glad that's what you took away as a highlight frommi.  I can tell you are keeping up with this stuff

[73 Reds]

49 minutes ago, gfp said:

 

Glad that's what you took away as a highlight frommi.  I can tell you are keeping up with this stuff

Me thinks @frommi's post was made in jest.  Can't AI distinguish sarcasm from reality?  If not, what good is it anyway? 

Edited February 11 by 73 Reds
missed line

[frommi]

1 hour ago, gfp said:

 

Glad that's what you took away as a highlight frommi.  I can tell you are keeping up with this stuff

I am too old and stupid to understand the other stuff, that was just a lot of blubber to me 

[Spooky]

Anyone starting to compile a database of materials that an AI tool can access and crawl? For instance, stuff like the Berkshire letters, Nomad letters, ML letters, PDF books, mental models / biases, etc.? We could build our own CoBF database / virtual investor.

[rogermunibond]

https://buffettbot.com/

[Castanza]

28 minutes ago, Spooky said:

Anyone starting to compile a database of materials that an AI tool can access and crawl? For instance, stuff like the Berkshire letters, Nomad letters, ML letters, PDF books, mental models / biases, etc.? We could build our own CoBF database / virtual investor.

 

For three years I kept an excel sheet which crawled the "What are you buying" thread and cataloged User, Stock and Price (USD) with 1k allocations per stock mention, per user, per day (If BRK.B gets mentioned 3 times that day then it gets allocated 3k). I treated options as stock purchases instead. Obviously far from perfect, but it was close enough to see a very rudimentary return of the users here both collectively and individually. Time period was sometime in 21->24. If I am remembering correctly the CAGR was something like ~18% collectively (don't quote me). 

 

Time for @Parsad to open a robo managed COBF ETF. 

[Spooky]

11 minutes ago, Castanza said:

If I am remembering correctly the CAGR was something like ~18% collectively

 

Pretty impressive result!

[Castanza]

1 hour ago, Spooky said:

 

Pretty impressive result!

 

Real results could vary widely though...I ignored sell side and position sizing etc. Plus the pricing was pulled through excel so some international stocks were not found etc. But I'd wager it's within ~2-5% of reality. 

[Spekulatius]

8 hours ago, Pelagic said:

For something like a law firm or medical practice where client/patient confidentiality are paramount and they're mainly handling a lot of documents that are primarily text or images local models make a lot of sense and in their current state are very capable at doing this. I don't know how many firms are racing out to have a local model installed on a cluster of mac-minis or something but price wise for ~$20k or so they'd be able to have Llama 3.1 running internally. Not inexpensive but also not unreasonable for most firms.

I think you can also purchase segregated cloud capacity at AWS/Google/Azure and run your data on a local model provided by the vendor and trained on your data if you have a business of reasonable size.

[formthirteen]

Quote

"We would note open-source software, which is free of licensing costs (essentially free development), has been available for customers to build their own applications for over 2 decades. Despite this, the market of third-party software has flourished in that time," Weiss added.

https://finance.yahoo.com/news/the-ai-fueled-software-meltdown-is-overblown-195456346.html

 

[NnnnotSoSmart]

AI’s Impact Accelerates
Feb 5, 2026
Companies globally that have been using AI for at least one year report double‑digit productivity gains as well as some workforce reductions.

 

https://www.morganstanley.com/insights/articles/ai-adoption-accelerates-survey-find?subscribed=true&dis=em_2026211_wm_5ideasarticle&et_mid=779936&et_mkid=&sfmc_id=252663634&new-parameter5=

 

 

 

 

AI Adoption Surges Driving Productivity Gains and Job Shifts | Morgan Stanley.pdf

[formthirteen]

18 hours ago, Castanza said:

For three years I kept an excel sheet which crawled the "What are you buying" thread and cataloged User, Stock and Price (USD) with 1k allocations per stock mention, per user, per day (If BRK.B gets mentioned 3 times that day then it gets allocated 3k). I treated options as stock purchases instead. Obviously far from perfect, but it was close enough to see a very rudimentary return of the users here both collectively and individually. Time period was sometime in 21->24. If I am remembering correctly the CAGR was something like ~18% collectively (don't quote me). 

 

Claude code was able to code this in ~15-30 minutes. It reached some interesting conclusions for (Feb 2025-Feb 2026) that I haven't fact checked:

 

Quote

The forum has an "anti-alpha" problem at the consensus level: the more people agree on an idea, the more priced-in it already is. The alpha sits in solo picks from skilled posters who bring differentiated research. This is consistent with the efficient markets view — the edge is in information asymmetry, not consensus.

 

Edited February 12 by formthirteen

[adventurer]

22 hours ago, gfp said:

This post on X is way too wordy but it might help break through some of the completely out of touch AI perspectives on this message board.  Maybe skip ahead a bit to save time.

 

There is an "aha" moment for AI skeptics that comes when you realize that the cutting edge AI tools (the ones that will be completely obsolete in 3 months) are better at doing your job than you are.  I've heard it referred to as the "Move 37 moment" - a reference to the AlphaGo documentary/competition, where Go champion Lee Sedol realized on 'move 37' that the AI was capable of creativity and far more advanced than he had given it credit for.  THAT WAS 2016!!!  Truly ancient history in this industry.

 

If you used a free version of ChatGPT 16 months ago for 15 minutes and think you know where AI is today you are who this article is aiming at.  

 

https://x.com/mattshumer_/article/2021256989876109403

Thank you. I needed that wake up call. I am 32 and technically not an illiterate. However this is something else...

Edited February 12 by adventurer

[Hektor]

14 hours ago, Spekulatius said:

I think you can also purchase segregated cloud capacity at AWS/Google/Azure and run your data on a local model provided by the vendor and trained on your data if you have a business of reasonable size.

Good point @Spekulatius. Will AI do the purchase/segregation etc?

[Castanza]

4 hours ago, formthirteen said:

 

Claude code was able to code this in ~15-30 minutes. It reached some interesting conclusions for (Feb 2025-Feb 2026) that I haven't fact checked:

 

 

 

What were the results? I also don't think this forum is "pricing-in" anything....Well maybe @gfp 

 

 

[formthirteen]

4 minutes ago, Castanza said:

What were the results? I also don't think this forum is "pricing-in" anything....Well maybe @gfp 

 

I didn't double check anything so don't take this analysis too seriously. The results were:
 

# COBF WAYB — Return Distribution

**Period:** Feb 2025 – Feb 2026

## Distribution Stats

| Metric            | Value  |
| ----------------- | ------ |
| Total allocations | 1,651  |
| Mean return       | +13.4% |
| Median return     | +4.1%  |
| Win rate          | 57.7%  |

## Return Buckets

| Bucket            | Count | % of Total |
| ----------------- | ----- | ---------- |
| > +100% (doubles) | 93    | 5.6%       |
| +50% to +100%     | 142   | 8.6%       |
| +20% to +50%      | 223   | 13.5%      |
| +0% to +20%       | 494   | 29.9%      |
| -20% to 0%        | 498   | 30.2%      |
| -50% to -20%      | 177   | 10.7%      |
| < -50%            | 24    | 1.5%       |

## Skewness

The distribution is positively skewed: the right tail (93 doubles) is much fatter than the left tail (24 picks down >50%). This is the classic value investing return profile — many small losses, occasional large wins. The 93 doubles contribute ~$93k to total P&L while the 24 big losers cost only ~$15k.

Mean (+13.4%) far exceeds median (+4.1%), confirming that a small number of big winners drive the aggregate return. Strip out the top 10% of picks and the portfolio return would likely be flat or negative.

## Implied Kelly

With a 57.7% win rate and roughly 2:1 payoff ratio on the tails, the Kelly criterion would suggest moderate position sizing. The crowd's equal-weighting approach ($1k per mention) actually works reasonably well for this distribution — it captures the right-tail upside without excessive concentration risk.

 

We still need humans to verify the output but that shouldn't take too long either with the help of AI. I just don't have the time right now.

[rogermunibond]

Been thinking about the longer term impacts of AI, if it is in fact a technological revolution, and not a bubble.

 

Past technological revolutions always mostly involved physical work and processes.

 

Agricultural revolution freed up labor, increased agricultural productivity, free labor moved into other higher value physical work. Mostly the increase in productivity fed into the industrial revolution, so goods that were previously handmade became industrialized. Each subsequent technology improvement fed into this. To the point where high tech manufacturing is relatively speaking incredibly labor efficient.

 

The next huge leap was the ICT technological revolution, which enabled huge amount of growth in services - software, accounting, legal, financial, engineering, etc. etc.  Knowledge work became more valuable, highly compensated activity, but did require considerable labor to produce outputs.

 

Now we have the next huge leap - the AI/agentic technological revolution. It promises to make knowledge industries more efficient, reduce labor requirements, and possibly result in price wars or increased margins. Maybe both simultaneously.

 

Where's the surplus labor going to go now? Maybe into services that can't be replaced by physical AI robots? Maybe into services that require a human premium? But what's that human premium and how impervious is it to improving AIs and physical AI robots?

 

 

[jfan]

https://www.darioamodei.com/

 

People may enjoy reading his thoughts here. One essay on AI benefits and the other on AI risks.

[Eldad]

14 minutes ago, rogermunibond said:

Been thinking about the longer term impacts of AI, if it is in fact a technological revolution, and not a bubble.

 

Past technological revolutions always mostly involved physical work and processes.

 

Agricultural revolution freed up labor, increased agricultural productivity, free labor moved into other higher value physical work. Mostly the increase in productivity fed into the industrial revolution, so goods that were previously handmade became industrialized. Each subsequent technology improvement fed into this. To the point where high tech manufacturing is relatively speaking incredibly labor efficient.

 

The next huge leap was the ICT technological revolution, which enabled huge amount of growth in services - software, accounting, legal, financial, engineering, etc. etc.  Knowledge work became more valuable, highly compensated activity, but did require considerable labor to produce outputs.

 

Now we have the next huge leap - the AI/agentic technological revolution. It promises to make knowledge industries more efficient, reduce labor requirements, and possibly result in price wars or increased margins. Maybe both simultaneously.

 

Where's the surplus labor going to go now? Maybe into services that can't be replaced by physical AI robots? Maybe into services that require a human premium? But what's that human premium and how impervious is it to improving AIs and physical AI robots?

 

 

Prostitution, cannon fodder, lots of new anti-technology terrorist groups with be recruiting. Maybe Bernie gets in charge and we go back to like a guild system for the workers. 

[rogermunibond]

Thanks @jfan

 

Amodei addresses this in The Adolescence of Technology, 4: Player Piano.  But alas, I don't think he does a very good job of discussing where humans may still be needed.  Rather he discusses more about how to allay the effects of this labor disruption. UBI anyone?

[NnnnotSoSmart]

One of the deeper dives on humanoid robots (physical AI) I've listened to (1Hr 43 min):

(I'd love to have one that could climb ladders and paint, clean windows, and hang Christmas lights)

 

 

Quick summary if you're short on time: 

https://metatrends.substack.com/p/figures-general-purpose-robot-why?utm_campaign=email-half-post&r=ag2qr&utm_source=substack&utm_medium=email

 

Edited February 12 by NnnnotSoSmart

[Valuebo]

Can someone explain me who will bear the support costs, liabilities, logistics, customer sourcing, ... of all these amazingly - and basically free - AI tools that every company will be using apparently in a few years? That the biggest companies in the world will do it all in-house over time, I can still understand and believe. But people are nuts if they think most companies' managers will run the career risk of dropping everything in some random AI developed tools on which everything has to rely. 

 

Hell, financial institutions are still running COBOL shit because everything is intertwined and risky to transfer. I'd love to see the avalanche of lawsuits that you have no way of guarding against when shit hits the fan because some AI fucked up.

 

Lots can change in 20-30 years time, but that is a given with or without AI. There is no way to predict whom will be the winners for that long a time frame but chances are many current incumbents will use AI tools to their advantage, as they have everything else up and running that is required to be a successful business in their field.

 

But what do I know. 10 years ago I also said true driverless cars are easily 20 years out. Even if it was fully here today, you have a shit ton of other problems to solve first anyhow. Meanwhile you have Elon Musk claiming it's 1 or 2 years out for more than a decade and that dude is the richest person in the world.

[frommi]

1 hour ago, jfan said:

https://www.darioamodei.com/

 

People may enjoy reading his thoughts here. One essay on AI benefits and the other on AI risks.

Thanks. He basically states that all human labor will eventually be replaced, no job is safe. If thats the end-game we will see massive deflation at first and the government taxing robot/compute time while giving universal basic income to everybody later on. At first you think companies like NVDA will profit from this long term, but i doubt it, because AI can just create a new factory and build the chips it needs itself. Same for everything. Maybe resources get more expensive, but AI could just figure alternative ways to build something or create energy. At that point humans will probably just there to serve and money doesnt matter anyway.