Martin Alderson

Theo video based on data from Tencent's Autocodebench: https://youtu.be/iV1EcfZSdCM Conclusion: Elixir was the best (had the highest problem solve rate). Reasons (Theo's interpretation): - code collocation, where documentation is integrated directly within the source code - design philosophy of a language (readability, clear idioms, and strict expectation management) --- Sources from video: - https://github.com/Tencent-Hunyuan/AutoCodeBenchmark/tree/main - https://martinalderson.com/posts/which-programming-languages-are-most-… - https://simonwillison.net/2026/Jan/19/nanolang/#atom-everything

Wrote this a while back. https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-cl… OpenRouter is the best guide to real costs.

ChatGPT’s revenue means nothing if reports are to be believed that it loses money on each paying customer on just inference. It’s definitely not enough to support its training costs. Sam Altman: We're very profitable on inference. https://simonwillison.net/2025/Aug/17/sam-altman/#:~:text=Subscribe,20… Independent analysis: Inference is very profitable. https://martinalderson.com/posts/are-openai-and-anthropic-really-losin… https://www.snellman.net/blog/archive/2025-06-02-llms-are-cheap/

> 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-p… > 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…

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…

It's well know everyone is making great money on inference. The cost is training. https://martinalderson.com/posts/are-openai-and-anthropic-really-losin… (with math working backwards from GPU capacity) "Most of what we're building out at this point is the inference [...] We're profitable on inference. If we didn't pay for training, we'd be a very profitable company" https://simonwillison.net/2025/Aug/17/sam-altman/ "There’s a bright spot, however. OpenAI has gotten more efficient at serving paying users: Its compute margin—the revenue left after subtracting the cost of running AI models for those customers—was roughly 70% in October, an increase from about 52% at the end of last year and roughly 35% in January 2024." https://archive.is/OqIny#selection-1279.0-1279.305 (Note this is after having to pay higher spot rates for compute because of higher than expected demand)

Exactly. I was wondering what language would be better for heavy agentic workflow. Due to a recent blog post, Clojure seem to be the most terse [1]. But some people are suggesting static types would help in this workflow. [1] : https://martinalderson.com/posts/which-programming-languages-are-most-…

The author doesn't seem to realize that a healthy margin has been built into the inference pricing. Once low cost open source inference providers get their hands on powerful frontier level models, there would be a severe margin compression for OpenAI and Anthropic. Source: https://martinalderson.com/posts/the-upcoming-ai-margin-collapse-part-…

> While we're in the token-equivalent of ZIRP, tokens don't cost what they cost, so there's sort of arbitrage to be had Do you have a source for this? I believe “personal subscription” plans on OpenAI and Anthropic are likely ran at a net loss or close to it, but all indications elsewhere are that API pricing for these companies and likely Google as well are profitable per API call [1][2]. I would definitely believe that the Chinese players are operating at a loss though if that is what you mean. [1] https://newsletter.semianalysis.com/p/anthropic-growth-and-bedrock-mix… [2] https://martinalderson.com/posts/are-openai-and-anthropic-really-losin…

If you want to invest into a company that barely does space at 10% and is 90% a datacenter Reits essentially at a P/E ratio of 90 whereas others have around 10 so you are gonna overpay for 10 times or a magnitude more Then sure, yes you are correct. https://martinalderson.com/posts/xais-new-rental-business/: seems to be a post on the frontpage of HN at the moment on what I am currently saying.

API prices of Anthropic, OpenAI, and Google are massively inflated. https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-cl… There's no way that all AI inference providers are colluding and/or all running at a massive loss, meaning the cheap Chinese model prices must be the real cost it takes to run frontier-class models PLUS their margin. Look at Deepseek 4 Pro. https://openrouter.ai/deepseek/deepseek-v4-pro/providers Deepseek and Baidu are subsidising prices but they probably train on inputs. I have no model training and ZDR in OpenRouter enabled, and the first provider that shows up there is Deepinfra, significantly more expensive than Deepseek. BUT much cheaper than Sonnet 4.6 and ChatGPT GPT-5.4.

https://martinalderson.com/posts/which-programming-languages-are-most-… This article claims that Clojure is the most token efficient language.

Original source please, that just links to: https://martinalderson.com/posts/open-weights-are-quietly-closing-up/ Which was submitted 10 hours ago.. perhaps upvote that submission? https://news.ycombinator.com/item?id=48036924

Various investigations have found it to be one of the most token efficient statically typed programming languages https://martinalderson.com/posts/which-programming-languages-are-most-…

13 Markdown files have erased 285 billion dollars of wealth in the stock market - https://martinalderson.com/posts/wall-street-lost-285-billion-because-… Why are you surprised that it takes only 65 lines of Markdown to create the next Linus Torvalds or Donald Knuth! I’m learning Haskell by Thinking before Coding, without working through the difficult exercises. I finally understand Monads because of Thinking first.

Remarkable that it echoes, from a different angle, this post from just a few days ago on HN: https://martinalderson.com/posts/has-the-cost-of-software-just-dropped… This last post was largely dismissed in the comments here on HN. Simon's experiment brings new ground for the argument.

Those are effectively made up numbers, since they're given to him by an anonymous source we have no way of corroborating, and we can't even see the documents themselves, and it contradicts not just OpenAI's official numbers, but first principles analyses of what the economics of inference should be[1] and the inference profit reports of other companies, as well as just an analysis of the inference market would suggest[2] [1]: https://martinalderson.com/posts/are-openai-and-anthropic-really-losin… https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSo… [2]: https://www.snellman.net/blog/archive/2025-06-02-llms-are-cheap/

https://martinalderson.com/posts/are-openai-and-anthropic-really-losin… All the labs are going hard on training and new GPUs. If we ever level off, they probably will be immensely profitable. Inference is cheap, training is expensive.

In short, the others have a huge margin if you ignore training costs. See https://martinalderson.com/posts/are-openai-and-anthropic-really-losin… for details.

"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-cl… 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-unleash… 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?

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-cl… 2: https://www.economist.com/finance-and-economics/2026/04/20/american-co…

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-sa… ² https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-cl…

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-cl… [2] https://martinalderson.com/posts/are-openai-and-anthropic-really-losin…