AI vs. software companies

[Spooky]

5 hours ago, Malmqky said:

Anyone wanna make a bet with me that OpenAI goes bankrupt?

You will have to give me some pretty good odds. OpenAI looks doomed.

[Spooky]

[tede02]

22 hours ago, Malmqky said:

Anyone wanna make a bet with me that OpenAI goes bankrupt?

 

I guess I wouldn't be totally shocked but that would be pretty wild and shake things up around the AI narrative!

[Malmqky]

2 hours ago, tede02 said:

 

I guess I wouldn't be totally shocked but that would be pretty wild and shake things up around the AI narrative!

 

Losing something like 14B per year

Guaranteeing private investors 17.5% returns aka can't raise funds for less

Cutting prices

Getting destroyed by Anthropic, Google etc.

Sam Altman is a wannabe Elon

Unsure how they will ever make money

 

They can probably avoid bankruptcy, but I'm very unimpressed with them.

[treasurehunt]

On 4/9/2026 at 12:51 PM, beerbaron said:

Does that Mythos thing remind you guys about when Japan was supposedly banning PS2 exports because it could be used in missile? Seems like a bit of marketing to me.

 

Did you read the paper on Mythos from the red team at Anthropic? I think someone had posted the link earlier, but here it is if you want to take a look: https://red.anthropic.com/2026/mythos-preview/

 

Seems like more than a bit of marketing to me.

[Whensthepaintdry?]

Does anyone have a good reason for why Amazon is investing in open AI

[roundball100]

18 hours ago, treasurehunt said:

 

Did you read the paper on Mythos from the red team at Anthropic? I think someone had posted the link earlier, but here it is if you want to take a look: https://red.anthropic.com/2026/mythos-preview/

 

Seems like more than a bit of marketing to me.

I agree, definitely beyond marketing.  The blog article gives convincing details on real security flaws.  So they are pushing boundaries.  On the other hand: if you look at the tremendous advances in the 13 years since the introduction of the AFL fuzzing tool by Michal Zalewski while at Google -- https://en.wikipedia.org/wiki/American_Fuzzy_Lop_(software) -- and improvements such as AFL++, and then throw in orders of magnitude more computing cycles as subsidized by the LLM craze, few software security experts would be surprised by these advances.  The ability to chain together multiple software exploits is not new, nor the ability to do this in a totally automated fashion (I can dig out references to such work at top security conferences 10-15 years ago already). But the continual pushing-ahead of tools that find new software vulnerabilities (so-called zero-days) is a worrisome thing.  If this interests you, a good read is Nicole Perlroth's 2021 bestseller, https://en.wikipedia.org/wiki/This_Is_How_They_Tell_Me_the_World_Ends

 

[treasurehunt]

1 hour ago, roundball100 said:

I agree, definitely beyond marketing.  The blog article gives convincing details on real security flaws.  So they are pushing boundaries.  On the other hand: if you look at the tremendous advances in the 13 years since the introduction of the AFL fuzzing tool by Michal Zalewski while at Google -- https://en.wikipedia.org/wiki/American_Fuzzy_Lop_(software) -- and improvements such as AFL++, and then throw in orders of magnitude more computing cycles as subsidized by the LLM craze, few software security experts would be surprised by these advances.  The ability to chain together multiple software exploits is not new, nor the ability to do this in a totally automated fashion (I can dig out references to such work at top security conferences 10-15 years ago already). But the continual pushing-ahead of tools that find new software vulnerabilities (so-called zero-days) is a worrisome thing.  If this interests you, a good read is Nicole Perlroth's 2021 bestseller, https://en.wikipedia.org/wiki/This_Is_How_They_Tell_Me_the_World_Ends

 

 

So Mythos is more of an incremental advance in capability and not a drastic improvement, in your view? Interesting. And thanks for the book recommendation; I'll try and check it out.

[roundball100]

18 hours ago, treasurehunt said:

 

So Mythos is more of an incremental advance in capability and not a drastic improvement, in your view? Interesting. And thanks for the book recommendation; I'll try and check it out.

Experts who have looked at the new Mythos tool and worked directly with it would be in a better position to say. From the link posted to the long blog, I'd call Mythos more than "incremental" (more than a minor advance to an existing approach), rather a next-step ... but not yet an instant game-changer. Nonetheless worrisome to the public (citizens of the world, versus investors per se) as it helps attackers more than defenders (for example, Perlroth discusses how today's major nation states have the capacity to take down each other's electrical grids by exploiting software vulnerabilities).  The reason for concern is that in our existing software-based world already (before Mythos), the ability to find software vulnerabilities already exceeds the time developers have to fix them ... new vulnerabilities are found at a rate higher than existing vulnerabilities can be fixed.  This imbalance has been growing over the past 10-20 years with advances in fuzz-testing tools (independent of AI).  The new Mythos tool makes this worse.  And, even if Mythos isn't openly released, it will be surprising if the general approach is not now replicated by other organizations/tools. 

[Spekulatius]

Interesting twist is that Anthropic is now designated a supply chain risk , so no Federal entity or contractor working for federal entities (including defense companies) can use it. Oh well.

https://www.cnbc.com/2026/04/08/anthropic-pentagon-court-ruling-supply-chain-risk.html

Edited April 17 by Spekulatius

[MungerWunger]

Well, IGV isn't down depiste a new claude coding model release:

 

https://www.anthropic.com/news/claude-opus-4-7

[gfp]

I tried out Opus 4.7 for the first time by asking it to speculate on the deal size of CSU's Derbysoft acquisition 

--------

 

I'll search for information about this acquisition to give you the most accurate estimate.The deal terms are confidential — the companies did not disclose financial terms of the transaction. So any estimate requires triangulating from what we know about Derbysoft and Constellation's typical acquisition pattern. Let me pull some data to build that estimate.Now I have enough to triangulate. Let me also check Constellation's typical acquisition profile.I have enough to build a credible estimate. Here's the triangulation:

Bottom line: likely ~$75M–$175M for the majority stake, with a central estimate around $100M–$130M

What we know

Deal terms were not disclosed. Constellation rarely discloses sizes on individual acquisitions unless they are material (typically >$100M), which itself is a useful signal — this one likely falls near or just under that threshold.

Target profile (Derbysoft)

• Revenue: Third-party estimates place Derbysoft at roughly $55M–$65M annually, with one source citing ~$61M and ~330 employees at ~$184K revenue/employee. Another trade-data source lists a broader "up to $500M" range, but that appears to be a category bucket, not an actual figure.
• Business: Hotel distribution connectivity (CRS/PMS-level), digital marketing, content management, and newer AI/flight modules. Classic vertical market software — exactly Constellation's strike zone.
• History: Founded 2002, only ~$15–30M raised historically, profitable enough to have never needed to IPO. Recently made its own tuck-in acquisitions (Arise, PKFARE), suggesting reasonable cash generation.

Constellation's acquisition math

• Constellation targets ~25%+ IRR hurdle rates, which in practice has meant paying roughly 1–3x revenue or 5–10x EBITDA for smaller VMS deals (Juniper Group specializes in smaller/mid-sized tuck-ins within Vela).
• Larger/distressed deals like Optimal Blue have gone at ~7x EBITDA; smaller proprietary deals often go higher on revenue multiples but are still disciplined.

Estimate build

Applying a typical Constellation range to ~$60M of revenue:

• Low case (~1.2x revenue): ~$70M enterprise value × ~60% majority stake ≈ $45M
• Base case (~2x revenue): ~$120M EV × ~70% stake ≈ $85M
• High case (~3x revenue, reflecting travel tech premium and AI/flight expansion): ~$180M EV × ~70% stake ≈ $125M

So a reasonable range for Juniper's check is roughly $75M–$150M, with ~$100M–$130M as the most defensible central estimate for the majority-stake purchase price. The enterprise value of the whole business is probably in the $120M–$200M range.

Caveats

• Travel tech commands a premium vs. Constellation's usual verticals, which could push the multiple higher than typical.
• Management retaining a minority stake is structurally important — Constellation often uses this to lower the upfront check and align incentives, which tends to compress disclosed headline numbers.
• The fact that Cantor Fitzgerald ran a process suggests Derbysoft likely attracted multiple bidders, which pressures the multiple upward.
• This is an estimate, not a leak. The actual number could land meaningfully outside this range; Constellation will only reveal it indirectly (if at all) via segment disclosures in future quarterlies.

[Spooky]

How is the new model 4.7? Claude has really been throttling my usage the last few days. 

 

For CSU to issue a press release it would need to be $100M+

[Valuebo]

Is there anyone who is of the mind that "AI" actually got smarter or at least edged more towards some form of determinism when comparing to say a year ago? It only seems to get better at things where it can lean on more training data and compute, like coding. I've held since years that actual AI is obviously nowhere near reality and I believe the market will grasp this in the coming quarters as well. It's gonna keep doing exceedingly well as a tool when used correctly, but at a financial cost and with the knowledge that you can't let it run freely without supervision or critical thought, as it can't ever actually "think" (despite model bullshit talk obviously claiming otherwise). I'm finally starting to see people around me wake up to the fact that these things just make shit up, lie, pull from untrustworthy sources and simply have serious limitations outside of some areas where it can be very powerful. Good to see this in actual day to day life and not just on Twitter.

 

As the saying goes, buy when the stock doesn't go down anymore on bad news. The story has hopefully reached max pessimism for software companies and new actual developments in the space (versus make believe and hype stories) are likely to taper off further, with each incremental gain to be smaller than the last and showing more and more limits and possible issues as time goes on. I'm really curious to see where we will find an equilibrium for usage versus cost, not only financial but also in terms of people and environment. Especially post subsidizing and peak AI delusions. 

[Spekulatius]

On 4/16/2026 at 10:38 AM, MungerWunger said:

Well, IGV isn't down depiste a new claude coding model release:

 

https://www.anthropic.com/news/claude-opus-4-7

The news are driving the share prices not the events.

[Spekulatius]

The next few years should be interesting for software cos. My guess is that  quite a few larger one will figure out how to save costs. software cos cost is  mostly wages and if it’s indeed true that developers can do far more with less than they can reduce the headcount by quite a bit without missing . DOmI think quite a few will show an immenese increase in margins.

 

SBC is an issue but not much as people think. The dilution rate is what matters most the GAAP expense and those rates have been coming down and are between 1-3% annually for most larger CO’s. Thats not much worse than many plain vanilla industrial or financial service cos.

 

Also, most SBC goes to new hires with annual vowing over 5 years or so. Much of the past stock grants are under water and with far less now.

 

Also, with the abundance of software engineers on the labor market- why do they need so pay the big stock grants upon hiring anyways?  i think some of the growers like CRM WDAY and others will soon turn into real cash machines. They are already have GAAP earnings and are FCF positive excluding SBC but this will become even more evident. Same for smaller cos like GTLB, HUBS and a few cybersecurity plays should be able to do the same.

Edited April 19 by Spekulatius

[frommi]

On 4/17/2026 at 2:27 PM, Valuebo said:

Is there anyone who is of the mind that "AI" actually got smarter or at least edged more towards some form of determinism when comparing to say a year ago?

There are tests for this, and yes it gets smarter but at least in the latest tests they are far away from humans. The problem with benchmarks is also that a lot can be fed back into the models as feedback and they are than just getting better in the benchmark but not in the real world.
But this is a test that is not pre-learnable because its about learning itself, the arc-agi-3 test:
 

Participant	Score (approx.)
Humans	100%
Gemini 3.1 Pro	0.4%
Gemma 4 (31B)	~0.30% (est.)
GPT-5.4	0.26%
Claude 4.6 Opus	0.5% And just a note: Gemma 4 is an open source model just released this month that runs on local hardware!

Edited April 19 by frommi

[Spekulatius]

Yes, LLM suck at things they have not been trained for, but are pretty good at things where plenty of training material exists. Programming is one where all the information is digitized and there is a lot of open source code available to train, so it’s ground zero for disruption. Others like healthcare also have enormous powntial but data is much harder to come buy.

 

I bet Epic will create great AI agents for all sorts of things in due time. they have a treasure drove of data but it’s most inaccessible in segregated systems and owned by their customers (or patients).

[whatstheofficerproblem]

Think RL is a fundamentally flawed way to model general intelligence. Problem here is you're pretty much domain locked to places where you have a tight feedback loop. The reason these models do so well in coding is because the feedback loop is tight and unambiguous i.e. either the code works or it didn't.

 

AI will 100% replace software engineers if judgement and management are zero value adds. The fundamental question that AI will replace software companies diverges into two different conversations. Will AI write most code? Almost certainly yes, and soon. The trajectory is clear we've gone from autocomplete to copilot to full agentic coding in about three years. Sundar himself said he started a project and didn't even know what language the AI chose until after it was running. Grunt work is disappearing.

 

Will AI replace the thinking behind software? That's much harder. What should we build? Why? For whom? How should systems interact? What tradeoffs are acceptable? What happens when requirements are ambiguous, contradictory, or politically loaded? That's the messy and there are no reward signals there. 

 

The likely outcome is something like what happened to farming. Agriculture went from 90% of employment to under 2%, but we produce more food than ever. Software output will explode while the number of people writing code by hand collapses. But someone still needs to decide what to grow and where to plant it.

 

The deeper irony is that AI replacing coders is actually bearish for the AI companies. If coding becomes near-free, the cost of building competitor products also becomes near-free. Every moat gets shallower. The same capability that lets OpenAI move fast lets three people in a garage replicate their product in a weekend. I don't mean that in a way someone will come up with GPT 5 overnight, but while the training moat is real, the product moat isn't. 

 

DeepSeek for example, while didn't replace OAI's infra, found ways to get comparable results with a fraction of the compute through architectural clevernessfound ways to get comparable results with a fraction of the compute through architectural cleverness. They didn't match the full stack but they matched the output well enough that it didn't matter. 

 

This pattern will become recurring imo. You won't need to replicate the infra but emulate the capability at the layer the customer touches. Three things make this possible over time. 

 

First is open-source models keep closing the gap. Sundar himself said it i.e. Gemma 4 is based on Gemini 3 architecture, fits on a USB stick, and the gap between it and the frontier is "both huge and not so huge in terms of time." The frontier leads, but the open-source tier follows faster each generation.

 

Second, distillation and fine-tuning are cheap. You can take an open-source base model, fine-tune it on a narrow domain, and get frontier-level performance for that specific use case at a tiny fraction of the cost. Most businesses don't need general intelligence and are content with AI that's great at their workflow.

 

Third, inference costs keep dropping. Even if you can't train the model, you can serve it. And if you're building a product on top of someone else's API, your competitor can build on the same API tomorrow. All of this yet again ties back to the fiber buildout parallels. The people who laid the fiber went bankrupt. The people who built applications on top of cheap fiber, the Google, Netflix, Amazons of the world, will come out on top. 

 

Companies like ServiceNow and GitLab have something the model providers don't i.e. embedded workflows and customer lock-in. ServiceNow doesn't win because it has the best AI. It wins because every Fortune 500 company has spent years wiring their IT service management, HR processes, and procurement workflows into it. Ripping that out is a nightmare regardless of how good a competitor's AI is. They can plug in whichever foundation model is cheapest or best at any given moment and pass the value to their customers. They're model-agnostic, which means they benefit from the competition between model providers driving prices down.

 

But there's a real threat to them too, in that if  coding becomes near-free, the cost of building a ServiceNow competitor drops dramatically. Today, replicating ServiceNow's platform would take hundreds of engineers years of work. In three years, maybe a small team with AI agents can build 80% of the functionality in months. The workflow lock-in is real, but it's not infinite, even more so if a customer is paying massive enterprise license fees and suddenly a lean alternative appears.

 

The real winners are companies that combine proprietary data with AI in ways nobody else can replicate. Think Bloomberg with financial data, or Epic with healthcare records. The model is commodity. The data isn't. That's where durable value lives.

 

The question then becomes, will the foundational model providers survive long enough for there to be budding competitors for incumbent software companies. I don't see how the model providers escape that trap.

 

Scenario one: they compete on price. Margins get crushed. They've spent billions on training and infrastructure, but the API becomes a commodity. The software companies love this as they get cheaper and cheaper inputs while maintaining their own pricing power through workflow lock-in. The model providers bleed cash, and the economics never work. Close enough, welcome back Telecom bubble.

 

Scenario two: they collude or consolidate to hold pricing and run the OPEC playbook. But unlike oil, you can't control supply here because open-source exists. The moment OpenAI and Anthropic try to maintain fat margins, Meta releases the next Llama, DeepSeek ships something competitive, and the software companies just switch. Once again, Sundar  Gemma 4 fits on a USB stick and is closing in on the frontier. The software companies being model-agnostic would make a pricing cartel impossible.

 

Scenario three: open-source wins outright. We go Linux trajectory & Linus Torvald didn't kill Microsoft or Oracle overnight, but he fundamentally restructured the economics of infrastructure software. If open-source models reach "good enough" for 90% of enterprise use cases and the trajectory here suggests they will, then the foundation model providers are selling the remaining 10% of premium capability at massive cost and that's a bad business to be in.

 

The really vicious dynamic is that all three scenarios can play out simultaneously across different market segments. Open-source eats the bottom. Price competition eats the middle. Only the absolute frontier commands premium pricing, but maintaining the frontier requires ever-increasing spend, and the lead time before open-source catches up keeps shrinking. You're running faster and faster to stay in the same place.

 

The one escape hatch seems to be vertical integration. Instead of selling the model, embed it in Search, YouTube, Cloud, Workspace, Waymo, and extract value through those products. The model becomes a cost center that accelerates your actual businesses, not a product you need to monetize directly. Microsoft is doing the same through Office, Azure, and GitHub.

 

OpenAI and Anthropic don't have that luxury. They are the model. Which is why OpenAI is frantically trying to become a consumer platform and Anthropic is pushing enterprise deeply. They need to build the application layer before the model layer commoditizes underneath them.

 

All this considered, software companies are probably in the strongest position of anyone in this chain. They're the arms dealers selling to both sides. The model providers are the ones with the existential math problem. And ironically, the better the technology gets, the worse the economics get for the people building it. Which is, once again, exactly what happened with telecom. So I think the software sell off makes no sense and the broader market is making the mistake of chasing the most exciting story instead of following the economics. 

 

Model companies have the breathless headlines, the frontier breakthroughs, the "this changes everything" demos. Software companies have become boring. Nobody writes breathless articles about ServiceNow's latest sprint. So it makes sense that capital flows toward the story. 

 

Another thing I think is being conflated here is that who captures the value versus who creates it. The model providers are creating enormous technological value but may capture very little of it economically, a la Telecom. The software companies are positioned to capture value by sitting on top of commoditizing infrastructure and maintaining pricing power through switching costs. The market is pricing model companies for the value they create and software companies for the disruption risk they face, when it should arguably be the reverse. 

 

Disruption risk is mispriced here in both directions. Software companies are being penalized for the threat that AI makes them replaceable. But as discussed above, their moat is workflow lock-in, customer data, and integration complexity, NOT CODE! Meanwhile model companies are being priced as if they'll maintain margin dominance in a market where open-source is closing the gap every six months and the product is inherently commoditizable. The market is overweighting disruption risk for the companies with moats and underweighting it for the companies without them.

 

As much as I hate bringing up ServiceNow as a comp here over and over again, they're running at 80%+ gross margins selling subscriptions with multi-year contracts all while model providers are burning cash at historic rates with no clear path to comparable margins because every efficiency gain gets competed away or replicated by open-source. The market is rotating out of proven high-margin recurring revenue into speculative negative-margin businesses where the competitive dynamics get worse as the technology improves.

 

To draw an non-Telecom parallel here, I would like the readers to ponder on this question for me. Circa 2006-2010, if you had to choose between investing in AWS itself or the SaaS companies that would be built on top of AWS, what would you choose? The right answer is the SaaS companies.

 

CRM, WDAY, NOW etc. rode the cloud commoditization wave and printed money because cheaper infrastructure made their businesses better, not worse. AWS also did well, but only because Amazon had the vertical integration escape hatch discussed above. The pure-play cloud providers that tried to compete with AWS on infrastructure alone mostly got crushed.

 

I am of the opinion that the market today is completely skipping asymmetry of outcomes. For a software company, AI going well means their product gets better and stickier. AI going badly means they keep running their existing business. For a model company, AI going well means they might monetize it before commoditization hits. AI going badly, hell, even going well but commoditizing would mean their entire business thesis collapses. The risk-reward is dramatically more favorable for the software names, but the market is pricing it backwards because the model company narrative is more exciting.

 

I have no crystal ball into when this might reverse, but I'd say OpenAI & Anthropic going public would be a start.

Edited April 20 by whatstheofficerproblem

[EgonKuhn]

That's a great summary @whatstheofficerproblem. Just one counter argument. Isn't the predicted AGI what sits in the drivers seat here trend wise? AI reaching far beyond the software world opening literally a new world of possibilities? Writing code being just the lowest hanging fruit because it's deterministic?

[whatstheofficerproblem]

11 minutes ago, EgonKuhn said:

Writing code being just the lowest hanging fruit because it's deterministic?

 

The counter argument defeats itself. Yes, you're right on it being lowest hanging fruit as it's deterministic. But even in this "easiest" domain, METR studies show experienced developers are 19% slower with AI, AI-generated code produces 1.7x more bugs, code churn has nearly doubled, and 95% of GenAI pilots fail to reach production. If AI can't reliably close the good-to-great gap in the domain with the tightest, most unambiguous feedback loop available, what exactly is the basis for believing it will handle domains where feedback is messier, more subjective, and more delayed?

 

The AGI argument has a dependency chain, current models improve -> we get AGI -> AGI solves everything. Each step here requires clearing a bar higher than the last. My point is that probabilistic models are the wrong mathematical tool for modeling reality. You will never reach AGI using RL.

 

Discussion around AGI feels a lot like cold fusion to me lmao. "sure, the current business model doesn't work, but once we achieve cold fusion, energy will be free." Prediction of AGI as you mention above, is not evidence of it but rather a bet. Who are people making the loudest predictions here? Usual suspects that are spending hundreds of billions into this bottomless pit with the most financial incentive to make those predictions.

 

I mean, Sundar himself dodged the AGI bullet and even called it semantics. Great, so even Google, spending $180 billion in capex, won't explicitly commit to an AGI timeline.

 

You'd have to assume a monotonic curve i.e. capabilities keep improving, therefore they will keep improving until AGI. But the research once again shows something radically different. Positive sentiment toward AI tools dropped from over 70% to roughly 60% in 2025. Trust in AI-generated code accuracy fell from 40% to 29%. AI projects have an 80% failure rate that has remained stubbornly consistent despite better tools and growing expertise. These are not the metrics of a technology on an exponential path to transcendence but looks like something hitting a ceiling and plateauing to me.

 

20 minutes ago, EgonKuhn said:

literally a new world of possibilities?

 

I allude to this in my write-up above. I'm not arguing AI isn't transformative, but we need to stop conflating technological potential with economic accrual. Fiber was important and transformative, who made money? Cloud was important and transformative, who made money?

 

The question never was if AI will be important, but who will make the money. Even if AGI arrived tomorrow, the economic questions we've been discussing don't disappear. They get worse. If AGI can do everything, then the model is the ultimate commodity as in everyone will have it, nobody can differentiate on it, and value accrues to whoever has the proprietary data, the embedded workflows, and the customer relationships. Which is, again, the software companies, not the model providers.

 

Think about what AGI actually means if you take the claim seriously. It means a system that can do any cognitive task at human level or above. Now follow that to its logical conclusion. If OpenAI builds AGI, what stops AGI from building a competitor to OpenAI? If the system is truly generally intelligent, it can design new architectures, optimize training runs, discover novel algorithms. The moment AGI exists, the intellectual moat a la "Saar we have the smartest engineers saar" evaporates because the AGI is the smartest researcher. And it's running on hardware that anyone with enough capital can rent. The scarcity was always human genius and doesn't AGI eliminate it?

 

The open-source dynamic accelerates this argument to absurdity. DeepSeek replicates frontier capability at a fraction of the cost using clever engineering. That clever engineering was done by humans who are scarce and expensive. In a post-AGI world, the clever engineering is done by the AGI itself, which means the replication cycle collapses from months to days or hours. Meta releases Avocado or Llama or whatever, AGI optimizes it overnight, and now everyone has frontier capability.

 

The model providers' lead time vs peers is their most valuable asset and moat and that would shrink to nothing. I said Google survives because it embeds AI and vertically integrates. But if AGI is generally available, every competitor can build equivalent products. The reason Google's integration works today is that it takes thousands of brilliant engineers years to build these systems. AGI removes that constraint and suddenly a startup can build a search engine, a video platform, a cloud infrastructure all in weeks. Google's moat wasn't really the AI, never was. It was the accumulated complexity that only a massive organization could manage. AGI is the universal solvent for accumulated complexity, at least that's what it inherently means.

 

Today, the stack looks like: hardware -> foundation model -> application -> customer. Value concentrates at the model layer because building models is hard. In a post-AGI world, building models is trivial as the AGI can do it. Building hardware is still constrained by physics as you need fabs, silicon, rare earth materials etc. So value either drops down to hardware and energy (things that are physically scarce and can't be replicated by intelligence alone) or rises up to the customer relationship layer (brand trust, regulatory approval, data access, contractual lock-in). The model layer, which is pure intelligence, becomes the cheapest part of the stack because intelligence is now abundant.

 

If you zoom out and look at all this retardation from a philosophical perspective, intelligence has never been the bottleneck for most economic activity. The bottleneck is trust, coordination, regulation, physical resources, and institutional legitimacy among other things. A brilliant doctor who isn't licensed can't practice medicine. A genius architect whose building isn't up to the building code gets shut down.  AGI adds infinite intelligence to a system where intelligence was already becoming abundant and other things were scarce. It's like adding infinite water to a desert where the constraint is actually arable soil.

 

So, AGI bull case is a paradox. The more powerful the intelligence becomes, the less the intelligence itself is worth, because it's the one thing that's no longer scarce. Everything else becomes relatively more valuable and what becomes valuable is not something these guys can provide.

 

That was all a mouthful, but yeah, been going down on this rabbit hole for while.

[formthirteen]

2 hours ago, whatstheofficerproblem said:

All this considered, software companies are probably in the strongest position of anyone in this chain. They're the arms dealers selling to both sides. The model providers are the ones with the existential math problem. And ironically, the better the technology gets, the worse the economics get for the people building it. Which is, once again, exactly what happened with telecom. So I think the software sell off makes no sense and the broader market is making the mistake of chasing the most exciting story instead of following the economics. 

 

Great post!

 

There's good news and bad news. AGI arrived two weeks ago:

... and tokens are a commodity, Jimmy tells me so:

https://chatjimmy.ai/

 

Generated in 0,014s • 15 760 tok/s 

[whatstheofficerproblem]

2 minutes ago, formthirteen said:

There's good news and bad news. AGI arrived two weeks ago:

 

That egg is constantly wrong about everything. All these guys including Jensen have been shilling "AGI has been here since yesterday" for 3 years now.

 

 

[EgonKuhn]

Great response @whatstheofficerproblem! What I meant was more that the AGI prediction seems to drive the whole "gold found in the desert story" here with writing code just being an example use case but it triggered you for a nice reply

 

Key is who will be the winner in the end economics wise and where will be the bottlenecks as you described. Great stuff so thank you very much!

[whatstheofficerproblem]

The problem with AGI is that once you have it you're close to singularity and once that happens world will be in an apocalypse with UBI, revolts and everything. Which is why it's all the more puzzling to me why anyone at all would allocate capital towards AGI, the ethical implications of it aside, it's literally impossible based on the current way we train models.

 

As for who wins, obvious answer would be Verisk, Constellation, Thompson Reuters & Epic systems kind of companies that have siloed prop data that these model sellers won't have access to. No amount of intelligence artificial or otherwise can conjure up 40 years of financial terminal data or the medical records of half of America's hospitals. The data is the alpha.

 

Then come the embedded workflow companies i.e. the ServiceNow, GitLab, Salesforces of the world. They win not because of what they build but because of what their customers have built into them. Decades of integration, customization, compliance configuration, and institutional muscle memory. They're model agnostic as discussed previously & they benefit from falling model costs, and their switching costs are measured in years of enterprise pain. AI makes their products better and stickier without threatening their position.

 

Then of course, AI infra names. The physical and the scarce is something AGI cannot replicate. All roads lead to TSMC, but you could go further down into HBM, NAND, Opticals, Energy, Cloud & neo-clouds etc.