louiereederson

Care to elaborate?

I know they acknowledge this but measuring autonomy by looking at task length of the 99.9th percentile of users is problematic. They should not be using the absolute extreme tail of usage as an indication of autonomy, it seems disingenuous. Does it measure capability, or just how extreme users use Claude? It just seems like data mining.

The fact that there is no clear trend in lower percentiles makes this more suspect to me.

If you want to control for user base evolution given the growth they've seen, look at the percentiles by cohort.

I actually come away from this questioning the METR work on autonomy.

You can see the trend for other percentiles at the bottom of this, which they link to in the blog post https://cdn.sanity.io/files/4zrzovbb/website/5b4158dc1afb211...

Your ability to sign up enough customers is directly related to your ability to sustain training costs. The model runs have a short lifespan. They may serve many customers at a given point in time per run, but in order to serve those customers over time you need to continually spend on training.

I think your mental model for an LLM vendor is similar to a foundry (i.e. TSMC). They spend a bunch of R&D on developing leading edge nodes and build foundries. That in your mental model would be similar to training costs.

My point is the correct mental model is more like (but not exactly like) a SaaS company, ironically. SaaS unit economics are a function of gross margin, churn and acquisition costs, i.e. Revenue x gross margin / churn - CAC. My point is some element (maybe the entirety) of training costs are more like CAC than they are like TSMC's R&D and capex. The question to ask to test this view is: is what happens to OpenAI or Anthropic revenue in 2027 or 2028 if they stop spending on training today? My view is it'll drop precipitously. This implies churn is very high. It is true that training costs can be spread over customers though, so the analogy breaks down there, but I think it is a better mental model than the foundry one.

Referring to my earlier comment, you need to have a model for how to account for training costs. If Anthropic stops training models now, what happens to their revenues and margins in 12 months?

There's a difference between running inference and running a frontier model company.

Anthropic reduced their gross margin forecast per external reporting (below) to 40%, and have exceeded internal forecasts on inference costs. This does not take into account amortized training costs which are substantial (well over 50% of revenue) and accounted for as occurring below gross profit. If you view training as a cost of staying in the game, then it is justifiable to view it as at least a partially variable cost that should be accounted for in gross margin, particularly given that the models stay on leading edge for only a few months. If that's the case then gross margins are probably minimal, maybe or negative.

https://www.theinformation.com/articles/anthropic-lowers-pro...

Big tech might ahead of the rest of the economy in this experiment. Microsoft grew headcount by ~3% from June 2022 to June 2025 while revenue grew by >40%. This is admittedly weak anecdata but my subjective experience is their products seem to be crumbling (GitHub problems around the Azure migration for instance), and worse than they even were before. We'll see how they handle hiring over the next few years and if that reveals anything.

You say this with such confidence and then ask if smaller chips require smaller wafers.

No. TSMC will not take the risk on allocating capacity to just anyone given the opportunity cost.

I think most estimates have Nvidia at more or less stable share of CoWoS capacity (around 60%), which is ~doubling in '26.

That is fair but doesn't do much to push back against the risk to the independent model vendors, particularly in consumer. This has represented a huge portion of the AI capex so far. OAI alone represented 2GW of compute in 2025. The point is they are in a fragile position, and so are the economics of AI DC spending in aggregate.

With respect to Google, I'd also wonder about the economics of AI search vs traditional.

At some point AI may deliver the level of net economic benefit you reference, but it's not entirely clear that we're there yet.

Right now much of the direct monetization occurs via OpenAI and Anthropic, who together have around $30B in annualized revenue. They are burning cash like crazy, though admittedly have potentially sustainable unit economics (gross margins around 40-60% before revenue share).

However, they need to spend a huge chunk of revenue on training. OpenAI spent something like $9b on training against around $13-14b in rev in 2025 (different from annualized rev) according to The Information. Anthropic's mix is supposed to be similar. Also implies a lot (maybe majority) of their compute spend is training.

If scaling laws falter, what happens to training spending? What happens to competitive degree of differentiation given Chinese open source models are a few months behind frontier? Then what happens to margins? It is very fragile.

I think these people are just as prone to behavioral biases as the rest of us. This is not a problem per se, it's just that it is difficult to interpret what is happening right now and what will happen, which creates an overreliance on the opinions of the few people closely involved. I'm sure given the pace of change and the perception that this is history-changing is impacting peoples' judgment. The unusual focus on their opinions can't be helping either. Ideally people are factoring this into their claims and predictions, but it doesn't seem like that's the case all the time.

To the point on Jevons Paradox, the number of people/developers joining GitHub had been accelerating as of the last Octoverse report. Related: "In 2023, GitHub crossed 100 million developers after nearly three years of growth from 50 million to 100 million. But the past year alone has rewritten that curve with our fastest absolute growth yet. Today, more than 180 million developers build on GitHub."

https://github.blog/news-insights/octoverse/octoverse-a-new-...

It is true that employment and pay are good. The question is a) are proper protections in place for workers and b) does pay adequately compensate for traumatic stress. Studying the issue is valuable, particularly given the potential for labor exploitation given the power imbalance between the largest, most powerful organizations in the world and a labor base which as you point out could otherwise be disenfranchised.

I think with this info we can assess whether the wage is indeed competitive. Otherwise claiming "well they're getting paid" can be used to justify exploitative labor practices or negate the need for ethical inquiry.

Also calling psychological harm metaphysical is questionable. Is PTSD experienced by American troops coming back from deployment metaphysical? Not necessarily comparing these cases in degree.

I'm not sure what invoking neocolonialism adds to the discussion. Best to engage on a factual rather than ideological basis.

This is partly a description of the concept of the business ontology. Bob Muglia has been discussing this as the central problem to be solved in the data platform space for at least a decade. You could view the battle between the various vendors vying for a role in enterprise AI as really a battle over who owns the business ontology or semantic layer.

This article does a good job explaining https://metadataweekly.substack.com/p/ontologies-context-gra...

This is a crowded solution space with participation from cloud, SaaS and data infrastructure vendors. All of these players and their customers have been trying to operationalize LLMs in enterprise workflows for 2+ years. Two big challenges are business ontology and fitting probabilistic tools into processes requiring deterministic outcomes. Overcoming these problems require significant systems integration and process engineering work. What does OpenAI have that makes them specifically capable of solving these problems over Azure, Databricks, Snowflake, etc., who have all been working on these problems for quite a while? I don't know if the press release really addresses any of this, which makes it seem more like marketing copy than anything else.

The question of lock-in is also a major one. Why tether your workflow automation platform to your LLM vendor when that may just be a component of the platform, especially when the pace of change in LLMs specifically is so rapid in almost every conceivable way. I think you'd far rather have an LLM-vendor neutral control plane and disaggregate the lock-in risk somewhat.

Yes you are missing the picture where Brandon asks Linguini to add a reminder to buy a pair of Arc'Teryx gloves, which Brandon is holding in his hands.

The point, as I noted below, is that this is an impractical solution.

You can justify the value of any ridiculous invention by comparing it to a world-changing invention.

The solution being discussed involves someone removing everything from their fridge, photographing it and paying for an LLM to process it into a database of sorts. Further, in order for this database to be complete they need to repeat this process every time something changes in their fridge. Also will the LLM be able to tell if my carton of milk is 10% empty? I do not disagree that food waste is a problem, but the solution seems laughably impractical, and the default (memory) seems far better suited to the task. I can confidently say that the net value creation is not comparable to the written word.

Do you have that much trouble remembering what is in your fridge to consider this the stupidest thing you have ever read on this site? I feel superhuman.

- Why do you need a reminder to buy gloves when you are holding them?

- Why do you need price trackers for airbnb? It is not a superliquid market with daily price swings.

- Cataloguing your fridge requires taking pictures of everything you add and remove which seems... tedious. Just remember what you have?

- Can you not prepare for the next day by opening your calendar?

- If you have reminders for everything (responding to texts, buying gloves, whatever else is not important to you), don't you just push the problem of notification overload to reminder overload? Maybe you can get clawdbot to remind you to check your reminders. Better yet, summarize them.

i don't know this author but ian bremner does this. it's as if he's conveying what he believes are serious and important thoughts in an unserious and casual way, to make it appear as if the thoughts - which again he probably thinks are brilliant - just come quickly and naturally. it comes across as performative though again not making claims against this author. and yes i am not using sentence case here, but this is not an essay.

i interpret figure 2 as showing that incoherence increases with model gens, albeit on a small sample size

that is really not the case in software, people commonly go between EV/S and EV/FCF for high growth names. also earnings could mean: GAAP earnings, non-GAAP earnings, ebitda, ebita, FCF, FCF ex share based comp.

you are aware of the long history of organizations being absolutely screwed by bad erp implementations right? nike's 2001 issue, the horrific birmingham oracle implementation, avon, etc.

jira premium is $15/mo/user for 300 users. you're saying $50k can cover developing the app inclusive of integrations, maintaining it, providing 24/7 service and 3 9s uptime (per the sla)? don't forget compliance and security. maybe the logic is everyone can be fired and replaced with agents?

>My hot take here is that the moats previously enjoyed by SaaS companies will increasingly vanish as smaller and smaller teams can assemble "good enough" solutions that companies will adopt instead of paying giant chunks of their budget on pre-built SaaS tools that will increasingly demand more training to Onboard.

why do people pay red hat/ibm for rhel? they earn pretty good margins too. to parent's point on software/=code

multiple of what? there is maybe one software company trading above 30x revenue - palantir. many companies growing at 20% trade at single digit revenue multiples.

this story is 2026 in a nutshell

i find the actual survey from PWC a bit ambiguous. they say 42% of respondents saw no impact from AI, but nothing about how much those respondents have invested in AI. what if they are only in POC, or have made miniscule investments?