Artificial Intelligence

[Milu]

2 hours ago, UK said:

 

Yes, this he i s very good. I think it is really good for some ego check, especially for all (including me) trying to bottom fish is software. I still do not have a strong conviction here, especially on the particular companies, but for me it is hard not to be afraid for some in the sector ending not that different from newspapers. valuation is not the most important thing here, past price performance is useless, I think it is all about survivability in AI era and so if one can answer this question, good, if not, this is gambling.

 

Regarding his materials and energy thesis, I have even less knowledgeable here comparing to software:), but can anyone please explain me, how all this can work if the mother of all commodities/energy, which is oil, is not up and also going parabolic? If you still gave cheap oil isn't it allows to still solve the majority of the problems with all the bottlenecks sooner or later?

I'd also encourage people to watch from minute 40 where he goes into Tesla and its potential edge in the self driving world we are entering into.

[UK]

https://archive.is/zXEKS

 

Why AI won’t wipe out white-collar jobs
The technology will expand their scope and raise their value

[nsx5200]

If you really want to get get a feel for AI capability on coding, you should be watching nerdy videos from coders about AI coding like this.  A TLDR summary is that AI coding is still a skill multiplier, and does not allow randos to create production-level products, which AI maximus believe.

 

I find that there's a huge gap between non-engineers and experienced engineers in their perception to create production quality products.  IMHO, new products are announced all the time in the media, so typical people think that with a few weeks/months, you can just have a product from start to finish.  Most don't hear about the failures unless somehow they go viral or do something extremely stupid.

 

While in between jobs, I used to work part-time for a company that helps people implement their random ideas for some compensation.  The amount of input resources that these people have in mind in order to realize the product is so fantastically low that it's not even funny.  Needless to say, without proper resource, most ideas failed, even if the idea is actually viable.  The company was also shady in that they would string the customer along just to suck their money without really telling them how unrealistic their expectation is.

 

13 hours ago, UK said:

https://archive.is/zXEKS

 

Why AI won’t wipe out white-collar jobs
The technology will expand their scope and raise their value

Companies see the short-term productivity gain, and they're letting the white-collar jobs go.  My humble long-term prediction is that, at the corporate level, this is going to turn out to be like automation, where all the competent companies will adopt AI.  So after it shakes out a bit, all the companies, relative to each other, are still going to be roughly the same, except more productive(especially if they can get rid of most management and overhead).  All the barriers to entry will be roughly the same, with more edge for the bigger corporations due to economy of scale in acquiring and deploying resources.  This will follow the winner-take-all trend that we've been seeing for decades now.  Whether the white-color jobs will increase(Jevon's paradox) or not (job obsolescence) looks to be context dependent, but most are still TBD.

[UK]

6 hours ago, nsx5200 said:

If you really want to get get a feel for AI capability on coding, you should be watching nerdy videos from coders about AI coding like this.  A TLDR summary is that AI coding is still a skill multiplier, and does not allow randos to create production-level products, which AI maximus believe.

 

I find that there's a huge gap between non-engineers and experienced engineers in their perception to create production quality products.  IMHO, new products are announced all the time in the media, so typical people think that with a few weeks/months, you can just have a product from start to finish.  Most don't hear about the failures unless somehow they go viral or do something extremely stupid.

 

While in between jobs, I used to work part-time for a company that helps people implement their random ideas for some compensation.  The amount of input resources that these people have in mind in order to realize the product is so fantastically low that it's not even funny.  Needless to say, without proper resource, most ideas failed, even if the idea is actually viable.  The company was also shady in that they would string the customer along just to suck their money without really telling them how unrealistic their expectation is.

 

Companies see the short-term productivity gain, and they're letting the white-collar jobs go.  My humble long-term prediction is that, at the corporate level, this is going to turn out to be like automation, where all the competent companies will adopt AI.  So after it shakes out a bit, all the companies, relative to each other, are still going to be roughly the same, except more productive(especially if they can get rid of most management and overhead).  All the barriers to entry will be roughly the same, with more edge for the bigger corporations due to economy of scale in acquiring and deploying resources.  This will follow the winner-take-all trend that we've been seeing for decades now.  Whether the white-color jobs will increase(Jevon's paradox) or not (job obsolescence) looks to be context dependent, but most are still TBD.

Thanks for your response!

 

Edited January 29 by UK

[Longnose]

7 hours ago, nsx5200 said:

If you really want to get get a feel for AI capability on coding, you should be watching nerdy videos from coders about AI coding like this.  A TLDR summary is that AI coding is still a skill multiplier, and does not allow randos to create production-level products, which AI maximus believe.

 

I find that there's a huge gap between non-engineers and experienced engineers in their perception to create production quality products.  IMHO, new products are announced all the time in the media, so typical people think that with a few weeks/months, you can just have a product from start to finish.  Most don't hear about the failures unless somehow they go viral or do something extremely stupid.

 

While in between jobs, I used to work part-time for a company that helps people implement their random ideas for some compensation.  The amount of input resources that these people have in mind in order to realize the product is so fantastically low that it's not even funny.  Needless to say, without proper resource, most ideas failed, even if the idea is actually viable.  The company was also shady in that they would string the customer along just to suck their money without really telling them how unrealistic their expectation is.

 

Companies see the short-term productivity gain, and they're letting the white-collar jobs go.  My humble long-term prediction is that, at the corporate level, this is going to turn out to be like automation, where all the competent companies will adopt AI.  So after it shakes out a bit, all the companies, relative to each other, are still going to be roughly the same, except more productive(especially if they can get rid of most management and overhead).  All the barriers to entry will be roughly the same, with more edge for the bigger corporations due to economy of scale in acquiring and deploying resources.  This will follow the winner-take-all trend that we've been seeing for decades now.  Whether the white-color jobs will increase(Jevon's paradox) or not (job obsolescence) looks to be context dependent, but most are still TBD.

I agree with most of this. 

 

Sure you can vibe code something that feels like a real project in a weekend.  However, to the point made above. To scale your vibe coded application unless you are an engineer and understand how to scale something your application would likely need to be torn down and recoded by someone who understands scalable dev architecture. The AI Vibe Coding will always struggle to project your future needs so they wont lay foundational things that allow you to bolt on or grow the project. Then if you dont know what you coded cuz the AI did it you wont know how to instruct the AI to fix it.  I am confident that again with a strong engineer prompting the engineer can read the generated code see the flaws an instruct AI to add foundational layers or bolt on points to the code to make things more scalable but it takes someone whos been there and done it to know what those are and where they are required. 

 

I agree on the automation / AI portion too. We are going to see some fun gains in efficiencies of companies that can figure out how to adopt well. 

[formthirteen]

A16z on AI:

 

 

 

 

Spoiler

The AI Application Opportunity: Key Investment Theses and Market Dynamics

Executive Summary

The current technology landscape is being reshaped by the Artificial Intelligence (AI) product cycle, a wave of innovation that builds upon all previous cycles—PC, internet, cloud, and mobile—to achieve an unprecedented speed and breadth of adoption. This briefing document synthesizes the core themes from a detailed analysis of this AI era, outlining the primary drivers of its growth and the key investment theses for identifying enduring companies.

The fundamental premise is that AI applications succeed by catering to a core aspect of human behavior: the desire to be "richer and lazier." By creating economic value while reducing labor, AI is unlocking immense opportunities, evidenced by the sharp inflection in enterprise spending and widespread consumer adoption.

Three primary investment theses emerge for AI applications:

1. AI-Native Software: New companies are emerging to challenge incumbents by building AI-native solutions for established software categories. Their primary path to market is through "Greenfield" opportunities—servicing new companies or those at a technological inflection point—rather than converting the existing "Brownfield" customer base of entrenched players.
2. Software Eating Labor: Representing what is arguably a much larger market, this theme involves creating entirely new software categories that perform jobs previously done exclusively by humans. These companies succeed not just by cutting costs but by generating significantly more value, such as increasing revenue or handling complex, multilingual, 24/7 tasks.
3. The "Walled Garden" of Proprietary Data: Companies that possess unique, proprietary, or hard-to-aggregate datasets are positioned to build powerful moats. By applying AI, they can transform this raw data from a low-value subscription product into a high-value "finished product," creating businesses that are exponentially more valuable and defensible.

In this new landscape, defensibility is more critical than ever. With the ease of creating software, durable moats are built not on features alone but on becoming an indispensable "system of record," owning an end-to-end workflow, or leveraging a compounding data advantage that is inaccessible to competitors, including the large AI labs. These principles apply equally across both the enterprise and consumer technology sectors.

--------------------------------------------------------------------------------

1. The AI Era: A Foundational Product Cycle

Product cycles have historically been the primary drivers of technological growth, from the PC and internet to the cloud and mobile eras. Each cycle consists of an infrastructure layer and an application layer that builds upon it, creating enduring companies like Microsoft, Amazon, and Shopify. The current AI era represents the latest major product cycle, but with distinct and powerful characteristics.

• Cumulative Foundation: Unlike previous cycles, AI does not start from scratch. It builds directly upon the foundations of cloud computing and the ubiquitous presence of smartphones—supercomputers in the pockets of a vast majority of the global population. This existing infrastructure enables immediate, widespread deployment and adoption.
• Unprecedented Adoption Speed: The adoption of AI technologies is occurring at a "break neck speed." For instance, approximately 15% of adults globally now use ChatGPT on a weekly basis, integrating it into their daily routines for a near-infinite number of use cases. This rapid integration is driving a massive share of net new revenue in the software industry.
• Rapid Innovation: The pace of progress is remarkable. The capabilities of AI models have advanced dramatically in just two years, moving from text and image generation to include native audio, real-time interaction, and sophisticated reasoning. The presenter notes, "it's hard to even imagine how far we've come even in just the two-year time frame."

2. The Core Driver of AI Adoption: "Richer and Lazier"

A central thesis is that AI's success is rooted in a fundamental view of human behavior: "everybody wants two things they want to be richer and lazier." They seek to do less work while gaining more economic value. Generative AI is the technology that directly unlocks this desire.

• Value Creation in the Enterprise: Contradicting reports of faltering enterprise AI deployments, expense data from companies like Ramp shows a "giant tick up" in spending. This indicates that tech-forward companies have moved beyond viewing AI as a "magic trick" and are now using it to save significant time and money, thus generating tremendous value.
• The Value/Cost Equation: AI fundamentally alters the economic calculation for tasks. Many valuable activities are not pursued because the cost of human labor exceeds the value generated. AI lowers the cost dramatically, making a vast number of previously uneconomical tasks viable and creating new markets.

3. Three Pillars of AI Application Investment

Three broad, defensible themes are identified as the most promising areas for building enduring AI application companies.

3.1 Theme 1: Traditional Software Goes AI-Native

This theme focuses on reimagining existing, well-defined software categories with AI at their core. This is analogous to how cloud-native companies like Salesforce and Shopify disrupted on-premise incumbents.

• The "Bingo Board": This metaphor refers to the grid of established software categories like ERP, CRM, customer support, and payroll.
• Greenfield vs. Brownfield Strategy:
  • Brownfield: Attacking an incumbent's existing customers (e.g., trying to replace an established NetSuite installation) is extremely difficult.
  • Greenfield: The primary opportunity for AI-native startups is to capture new customers. This includes brand-new companies setting up their software stack for the first time or existing companies that have hit an "inflection point" (e.g., outgrowing QuickBooks) and are evaluating new systems.
• The Power of Incumbents: It is a mistake to assume incumbents will be defeated. Companies like Workday, SAP, and Adobe will integrate AI to make their products stronger and create new revenue streams, leveraging their existing locked-in customer bases.
• Moat: The System of Record: The most durable companies in these categories become the "system of record," meaning the business cannot operate without them. This creates immense stickiness. A key phrase used to describe this dynamic is, "the best companies have hostages not customers."

3.2 Theme 2: New Category Creation - Software Eating Labor

This is considered the largest potential opportunity, as it involves creating software solutions for jobs previously performed by people, a market astronomically larger than the existing software market.

• The Governing Principle: Instead of competing on a software "bingo board," these companies look at job boards. A job posting, such as for a front-desk receptionist, becomes a feature list for a new software product that can perform a majority of the listed responsibilities.
• Value Proposition: The focus is often less on cost savings and more on superior performance. AI software can work 24/7, speak multiple languages, and maintain perfect compliance, generating value far beyond what a human counterpart could.
• Case Study: Eve (Legal AI)
  • Market: Serves plaintiff-side attorneys who work on contingency, meaning they only get paid if they win.
  • AI Alignment: This business model creates "incredible alignment" with AI. Making attorneys more productive directly translates into higher revenue, as they can take on more winning cases.
  • Workflow & Moat: Eve aims to own the entire end-to-end workflow, from client intake to case outcomes. This generates a proprietary dataset on which cases are valuable. This data is not public and cannot be scraped by large labs, creating a powerful, compounding competitive advantage. The more cases it processes, the smarter its intake model becomes.
• Case Study: Salient (Auto-Loan Servicing)
  • Value Generation: The key selling point is not replacing a $50 million call center but that its AI agents collect 50% more revenue.
  • Superior Capabilities: The software masters the complex web of collection laws across all 50 states and various counties, something no human team can do perfectly. It also operates without the high employee churn common in these roles.
  • Data Moat: The company's defensibility comes from its data on what collection scripts work. Having conducted millions of calls, it knows "exactly what to say," making it extraordinarily difficult for a competitor to replicate its success.

3.3 Theme 3: The Walled Garden - Proprietary Data as a Moat

This theme centers on companies that control a unique dataset. AI transforms this asset from a simple commodity into the foundation for a highly valuable and defensible "finished product."

• The Metaphor: Large AI labs (like OpenAI) are compared to a vegetable farm selling raw ingredients (tokens). A "Walled Garden" company is a restaurant that uses those ingredients but also has its own exclusive, unique vegetables to create a finished meal that the farm cannot replicate.
• From Data to Product: Before AI, companies like Pitchbook or FlightAware sold subscriptions to raw data. Now, they can use AI to deliver a finished product the customer actually wants (e.g., an analyst memo comparing a startup to its competitors), capturing significantly more value.
• Sources of Proprietary Data: The data does not need to be created; it can be aggregated from public but fragmented sources. The moat is often built over time by collecting information that is free today but becomes proprietary as a historical record (e.g., the number of YouTube subscribers a channel had in 2017).
• Case Studies:
  • Open Evidence: Built a ChatGPT-like interface for doctors, but its power comes from an exclusive license to the New England Journal of Medicine and other premier medical journals.
  • VLEX: A 26-year-old company that digitized Spanish legal records. Its revenue quintupled after adding an AI layer that could produce finished legal memos incorporating its unique data.
  • Ask Leo: A procurement product whose moat is a proprietary collection of old contracts from major vendors, allowing it to advise clients on negotiation points that are not public knowledge.

4. Application in the Consumer Sphere

The three investment pillars identified in the enterprise market are mirrored directly in consumer AI.

Theme	Description	Consumer Example
Traditional Goes AI-Native	An AI-first product emerges to challenge an established category incumbent.	Korea: An AI-native design tool positioned as the modern alternative to Photoshop for a new generation of creators.
New Category Creation	A new market is created around a novel AI capability.	11 Labs: Created a market for high-quality voice and audio models for both consumer and enterprise use cases.
Proprietary Data	A consumer product is built on a unique, defensible dataset.	Slingshot: An AI therapist trained on a proprietary foundation model built from anonymized notes taken by its AI scribe during real therapy sessions.

Additionally, a key strategy for consumer startups to compete with Big Tech is the Aggregator Model. Like Kayak, which aggregates flights from all airlines, these startups provide a single interface to access the best models from all labs for a specific task (e.g., creative tools). This is a position the labs cannot occupy, as they are incentivized to promote only their own models.

5. Moats and Defensibility in the AI Era

The rapid pace of development and the ease with which software can be created ("vibe code") means that durable competitive advantages, or moats, are more important than ever. The analysis distinguishes between simple differentiation and true defensibility.

• Differentiation: A unique feature, such as a novel AI capability. This is often temporary and can be replicated.
• Defensibility: A structural advantage that is difficult to overcome. Key sources of defensibility in the AI era include:
  1. Workflow Ownership: Becoming the system of record or owning the entire end-to-end workflow for a critical business process (e.g., Eve).
  2. Compounding Data Advantage: Creating a feedback loop where product usage generates proprietary data that improves the product, making it smarter and stickier over time (e.g., Salient, Eve).
  3. Walled Garden Data: Owning an exclusive and valuable dataset that cannot be accessed by competitors or large AI models (e.g., VLEX, Open Evidence).

 

[MungerWunger]

https://basehitinvesting.substack.com/p/thoughts-on-market-valuations-ai

[MungerWunger]

 

[gfp]

I found this introduction to Claude for Excel useful.  Claude / Anthropic has made deals for a lot of very handy proprietary data feeds (LSEG, S&P Capital IQ, Daloopa/SEC Edgar, earnings call transcripts, Moody's, Morningstar, Pitchbook, ThirdBridge for expert network interviews, etc) that are particular useful to people in our line of work.  Made available to Claude pro subscribers on Friday according to this guy

 

https://www.youtube.com/watch?v=f-v0fJgBqhk

 

(youtube link to a video where a dude in a Seahawks jersey explains a bit about the newly widely available Claude for Excel)

 

If you search around youtube there are several intros to this including this guy's:

https://www.youtube.com/watch?v=m5YF89Kym1Y

(another youtube link)

 

...

Plus some pre-built "agentic skills" that do Jr. Analyst type tasks out of the box:

 

 

Gemini suggested I might want to use these data feed connectors that Anthropic has contracted for in ways like these:

 

Edited January 31 by gfp

[formthirteen]

These videos help explain the current state of AI, including M&A outlook, from the viewpoint of Silicon Valley. Sounds like Wall Street is listening... Value investors, not so much...

 

M&A outlook:

- "Whatever the question is, the answer is maybe, it used to be no"

- "This year could be the biggest M&A year in history"

[Luke]

 

 

Actually insane how good this is...

[KJP]

On 1/31/2026 at 12:18 PM, gfp said:

I found this introduction to Claude for Excel useful.  Claude / Anthropic has made deals for a lot of very handy proprietary data feeds (LSEG, S&P Capital IQ, Daloopa/SEC Edgar, earnings call transcripts, Moody's, Morningstar, Pitchbook, ThirdBridge for expert network interviews, etc) that are particular useful to people in our line of work.  Made available to Claude pro subscribers on Friday according to this guy

 

https://www.youtube.com/watch?v=f-v0fJgBqhk

 

(youtube link to a video where a dude in a Seahawks jersey explains a bit about the newly widely available Claude for Excel)

 

If you search around youtube there are several intros to this including this guy's:

https://www.youtube.com/watch?v=m5YF89Kym1Y

(another youtube link)

 

...

Plus some pre-built "agentic skills" that do Jr. Analyst type tasks out of the box:

 

 

Gemini suggested I might want to use these data feed connectors that Anthropic has contracted for in ways like these:

 

 

Thanks for the heads up on this.  Have you tried it yet?  

[gfp]

Just now, KJP said:

 

Thanks for the heads up on this.  Have you tried it yet?  

 

No I haven't yet.  I had a Claude pro subscription to try out Cowork and used it to do some very token-intensive work on my desktop and it did a very complicated task correctly, exactly as I specified.  But I also continually ran up against token limits and had to wait and finish after a few hours.  So I asked Claude and Gemini how to get my $220 back and they said "do this" so I went and got my $220 back.

 

Then a week later they come out with Claude for Excel and I wished I hadn't cancelled it.  So I will probably pony up the big bucks again.  I'm not a huge spreadsheet junkie but I do use Excel often and I'm not very fast at it so this tool would be helpful for me personally.

[gfp]

This video is a couple weeks old now (Davos) but I watched it last weekend and think it is pretty great to hear from two of the leading minds in AI on stage at the same time with a competent interviewer.

 

Demis Hassabis and Dario Amodei are both primary sources worth listening to - plus Dario's resting facial expressions are endlessly entertaining.  

 

https://www.youtube.com/watch?v=9Zz2KrBDXUo

 

[formthirteen]

Good news and bad news:

http://finance.yahoo.com/news/why-the-ai-driven-software-stock-crash-probably-isnt-over-125809330.html

 

Quote

"There's this notion that the tool in the software industry is in decline, and will be replaced by AI ... It is the most illogical thing in the world, and time will ‌prove itself," Huang reportedly said at an AI event that router giant Cisco (CSCO) held on Tuesday. "If you were a human or robot, artificial, general robotics, would you use tools or reinvent tools? The answer, obviously, is to use tools ... That's why the latest breakthroughs in AI are about tool use, because the tools are designed to be explicit."

 

Once the market understands this we might see a change in sentiment. It's like the PC-to-Cloud migration, some companies were able to do it (ADBE, TEAM, etc.), others are not so good at adapting (PYPL). It might take a quarter or two for a change in sentiment, or it could be tomorrow, checks stock prices, any second. Software = a tool ("free as in beer"). Software created and audited by a bunch of humans and sold by an enterprise = an enterprise tool ($-$$$$).

[Spekulatius]

TEAMS, WDAY, CRM and another are systems of record. Messing with them means potentially corrupting stored data and disrupting  workflows with a lot‘s of users touchpoint. It‘s not going to happen that easily. There are also data and cyber security concerns with replacing them.

 

I think it might actually be easier to go heavy on AI with chip design given that all the chip designer cos have large libraries to train on.

[Spekulatius]

Google to OpenAI after announcing $175B in Capex for 2026:

[Milu]

10 minutes ago, Spekulatius said:

Google to OpenAI after announcing $175B in Capex for 2026:

Yes looks to be google playing to its strength in that it has a business expected to earn 150-200 billion in operating cash flow and they are planning to spend it all, open AI has to rely on capital raises which is. It as straightforward. We are getting into uncharted territory here with the level of investment so hopefully they are seeing the returns or pathway to returns. 

Edited February 5 by Milu

[NnnnotSoSmart]

48 minutes ago, Spekulatius said:

Google to OpenAI after announcing $175B in Capex for 2026:

Damn the torpedoes...

[Spekulatius]

1 hour ago, Milu said:

Yes looks to be google playing to its strength in that it has a business expected to earn 150-200 billion in operating cash flow and they are planning to spend it all, open AI has to rely on capital raises which is. It as straightforward. We are getting into uncharted territory here with the level of investment so hopefully they are seeing the returns or pathway to returns. 

OpenAI is painfully slow compared to Gemini

and Gemini‘s output is about as good. I want my AI slob fast .

 

Maybe OpenAi is toast. Claude is slow too but has their strong suit in coding etc.

[MungerWunger]

On 10/3/2025 at 8:07 AM, billybobjovialdechicoutimi said:

Maybe it's been discussed before... pardon me if so and thanks in advance for a link to the relevant thread... but

 

How is it that the market is terrified that CSU's business will be crushed by AI, but nobody has such fears for Shopify... (or at least, its stock only gets an erotic caress)

 

Can AI not reproduce in one second Shopify's entire infrastructure?

 

Thanks for any insights, I am not smart enough to get it myself

Seems you were just early Billy ... It's also being taken to the woodshed now lol

[billybobjovialdechicoutimi]

1 hour ago, MungerWunger said:

Seems you were just early Billy ... It's also being taken to the woodshed now lol

thanks for the ref @MungerWunger

 

[NnnnotSoSmart]

Orbital Data Centers

 

• Elon: "In 36 months, the cheapest place to put AI will be in space:

 

Discussion starts at 00:00 min.  The entire interview is almost 3 hrs long.  You may want to grab a beer (or two?) before sitting down to watch.

 

 

Edited February 5 by NnnnotSoSmart

[zzzyx]

On 1/31/2026 at 9:18 AM, gfp said:

(youtube link to a video where a dude in a Seahawks jersey explains a bit about the newly widely available Claude for Excel)

Thanks for that link - really good info (though the dude is bit wordy - could have got all the points across in half the tiime)

[gfp]

Always worth a read

https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point?utm_source=post-email-title&publication_id=6349492&post_id=186989157&utm_campaign=email-post-title&isFreemail=true&r=oqhe5&triedRedirect=true&utm_medium=email