Martin Alderson writes about AI economics and developer tooling, including inference costs, token efficiency of languages and frameworks, and model speed.
HN commenters cite it on whether AI labs lose money on inference, using it as a reference on Claude Code costs and language token use.
Equating Clojure and Python just because they both dynamically typed is a very shallow comparison. The actual, practical experience is like a day and night. We have projects in Python, Clojure, Java, C# and Golang. Clojure is by far the least problematic to deal with. Even though they have many ugly and non-idiomatic, very old parts.
LLMs, matter of fact do work beautifully with Clojure, specifically because of the "true" REPL. Python doesn't have a comparable REPL - at best, Python's is an interactive shell.
When you give an LLM a closed loop system where it can evaluate code in a live REPL and immediately observe the results, it stops guessing and starts reasoning empirically.
With Clojurescript, you get a REPL connected to a browser - LLM can navigate any element, click buttons, have the entire page context to inspect and alter - all that without any compilation, without even saving and reloading any code anywhere. It seems you have zero idea how amazingly liberating the actual experience that is. And btw, it's the most token efficient language¹.
> ClojureScript is not a major language
a) There's plenty of Clojure code in the wild now and it's been there even before we started training LLMs.
b) The language is very small syntactically - it gives LLMs fewer ways to fantasize some weird constructs.
c) More popular languages have dozens of ways of doing similar things. Clojure's community is smaller and organized, there's less fragmentation - they don't have twenty…
> Why would there be stolen credentials in a sandbox?
The claim is not that there were stolen credentials in the sandbox - it's that the agent found a way OUT of the sandbox and gained access to the public internet, then obtained the credentials from there.
> This means the model had the ability to build resilient infra and open bidirectional ports. Create and deploy scripts/apps so it can maintain state across multiple VMs?
Yes. If you've spent much time with Fable 5 or GPT-5.6 Sol you shouldn't find that particularly surprising. These models can get a LOT done.
> I'd be curious to find out what kind of sandbox
My best guess right now is a container with networking disabled except for the IP address of a dedicated HTTP proxy server which was running a proxy that only allowed access to PyPI and a few other packaging sites.
I wrote about how OpenAI's production version of that worked (based on Artifactory) back in January: https://simonwillison.net/2026/Jan/26/chatgpt-containers/#installing-packages-from-pip-and-npm
> Hugging Face It's not an epitome of cyber security
This story put that well: https://martinalderson.com/posts/huggingface-openai-exploit/
> A final point on this - Hugging Face has an enormous attack surface. They have more interfaces than I can count which run untrusted models and code. While they definitely have invested in defences, by nature of their operating model they do have many more opportunities to be attacked than many other services. I…
"The biggest thing we’ve learned from the large language model generation is how many people are excited to replace human beings, and how many people just don’t understand labour of any kind"
Really? I think we've heard C-suite blabbermouths and borderline nontechnical tech CEOs being amplified by social media making these assertions, but boy did the weakly efficient market deliver a relatively swift correction to that mindset, no?
Q: "What would it take to change your point of view?"
A: "[AI] would have to solve all hallucinations forever, which they are completely incapable [of]."
Can you produce a human that is infallible? I'll wait.
Finally: https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-claude-code-user/
My hot take is AI is increasingly less unprofitable as the cost of serving tokens drops and Nvidia's ongoing offers to guarantee profitability is a sign that it isn't stopping anytime soon.
https://newsletter.semianalysis.com/p/nvidia-gpu-debt-backstop-unleashes
You either believe in the underlying science and technology or you don't. But in a world where AI is a fad like cabbage patch kids and beanie babies, what's next?
Cost to the business and price to customers are not the same thing. Even if it cost Anthropic nothing to run any of these (it actually costs quite a bit in electricity, infrastructure, ops teams to keep everything running smoothly, and above all else, extraordinary R&D expenses to develop the models), they could set the price at a million dollars per token if they wanted to.
That clarified: yes, every major lab is losing money on full utilization of their inference subscription plans. The API prices are what the business has determined they need to achieve profitability, and are not reflective of actual costs as you point out, but the discounts vs API pricing can get pretty extreme. Some users report 50x+ (98%+) discounts on the $100/mo Max subscription plans vs PAYG API pricing¹. Even the skeptical, contrarian takes that focus on cost to the business will tell you that, yes, Anthropic is losing money on those subscriptions, even using generously low estimates on costs².
¹ https://www.ksred.com/claude-code-pricing-guide-which-plan-actually-saves-you-money/
² https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-claude-code-user/
This video is unconvincing. The thesis of the video basically boils down to "AI companies are using electricity, driving up prices, and residential consumers are paying for it". However it neglects that most of the power usage is caused by the same residential users, who are getting something out of it. 80–90% of AI energy usage is estimated for inference (eg. generating responses on chatgpt), not training[1]. Moreover despite AI companies being unprofitable as a whole, they're making fat margins on inference[2]. Therefore it's reasonable to assume that most of AI electricity consumption that the video complains about is as a result of ordinary people using AI. Taking this into account, blaming AI companies for jacking up electricity prices in this context makes as much sense as blaming airlines for jacking up oil prices. They're only doing so because people are buying their services, and presumably deriving some sort of utility from it.
[1] https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
[2] https://martinalderson.com/posts/are-openai-and-anthropic-really-losing-money-on-inference/
I don’t necessarily disagree but to provide some counter points:
1. Model providers are currently profitable when just counting the cost to serve tokens for inference[1]. They lose money training the next generation of models.
2. Open models don’t work nearly as well. Given that tokens are still relatively cheap, and hallucinations are expensive, I’ve not seen a huge up tick in open model usage for coding agents yet.
3. On the AI economy front, I really have no idea, but AI companies (meta, msft) have already come down in value. It seems investors are at least a little wary of AI over valuation. Of course, the stock market is not the economy, but it’s not clear where warning signs would be. Earnings are healthy.
1: https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-claude-code-user/
2: https://www.economist.com/finance-and-economics/2026/04/20/american-corporate-profits-keep-shrugging-off-global-tumult