721–730 of 1001 posts
Maybe acceleration in smaller teams. We still seem in the era of the early internet where what questions LLMs change hasn't exactly emerged.
I know everything you’ve done for the tech community, but I please you to take some time off and reflect on this article. It’s not on par with ur usual level, but the tendency has been visible from the last couple of articles.
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down. This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer. That's a _huge_ shift. Most people…
I will also tell you, as someone who works at a company that's trying to remain profitable, that token spend has caught the eyes of finance and much like cloud spend they've already started applying pressure to control costs. This May my team is protected to use 30% fewer tokens than we did in April - this was by intention. I suspect we'll drop more in June.
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down. This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer. That's a _huge_ shift. Most people…
I work for a tiny little company ($150MM annual rev with 9% net) and we are already looking at dropping $100k on hardware to run local models because, for us, they're "good enough." Our estimated spend for AIaaS would exceed that cost in less than a year. In a few years, there will be hardware capable of running frontier models good enough for most things at accessible prices for even tiny companies.
Unfortunately Simon drank from the AI Cool Aid. I know everything you’ve done for the tech community, but I please you to take some time off and reflect on this article. It’s not on par with ur usual level, but the tendency has been visible from the last couple of articles.
I thought this was one of my best pieces of writing this year.
(In case you missed it, the title was meant as a subtle burn on those two companies - it's pretty absurd for them to only just be finding product-market fit when they're already valued at over a trillion dollars.)
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down. This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer. That's a _huge_ shift. Most people…
Maybe it's just me being (trigger warning from me providing an honest self assessment) very intelligent + a generalist, but i went from only full stack webdev and .NET to being able to implement an end-to-end LLM training pipeline (data curation, tokenizer, pretrain, sft, DPO - using ~$100 in cloud compute to train a class-competitive 1B STEM model)...and a full economic financial modeling and quant analysis application that pulls up to date economic, economic, news, stock data from the entire world and uses Dagster to orchestrate tech ical indicators and fundamentals and signals... and i did these things for learning and for fun. i built my own sublime text and obsidian replacement. i built my own reddit/twitter/hackernews/substack/news aggregator. i built countless other useful tools and utilities for me personally and for work I build more that empowers multiple departments.
Ive built 2 browser games, one already released to great reviews and 100k+ hours played. Ive built a tool on top of claude code that does ~60% of my job. Ive run data analysis on company financials for forecasting that have been refined and are producing very accurate predictions. Ive built competitive analysis tools and trackers.
All of this in 3 years. The projects are all clean, documented, with great code practices and modularity. A purist would surely consider some of the code slop. But it all works completely and fills real needs.
This is a huge shift. Anyone not realizing it yet is just simply behind the curve. I would not have accomplished 1/10 of this without AI coding. I went from copying code into and out of browser chats for 2 years before getting on the CLI train, and it is absolutely ridiculous the ROI you get from subscriptions to Claude or Codex.
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper? I don't see the business model working. My closest friend actually does automation software for large companies. He does not use Claude or openai at all. He primarily uses gpt 120b on cerebras and glm-5.1 for heavy thinking work. And some other small models for various tasks. All open source. And thes…
GLM-5.1 isn't just as good. It is no match for Opus running in Claude Code. Please try it yourself. Open source models are about a year behind at least.
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down. This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer. That's a _huge_ shift. Most people…
Here are a few thoughts: - The publicly available information about how inference costs compare to training costs is conflicted. EEs involved in datacenters talk about power usage spikes during training runs as if they were a major factor in the designs, but academic papers discussing cost-optimal scaling confidently treat inference-time compute as a major factor. - On the side of the balance indicating that training…
Training involves multiple passes over the entire training dataset, ideally in large batches where you can perform inference on as many samples as possible simultaneously and then perform backpropagation to adjust the model weights (which is about as expensive as inference).
Let's consider the size of the dataset we're dealing with here. The dataset likely consists of practically every piece of digitized text they can get their hands on (including that extracted from audio and video). We know Google has digitized a large portion of the books in existence as part of their "search book contents" feature and we have no reason to believe they're not using it alongside their cache of 90+% of the internet to train their models. We're talking about 100s of millions of books each with an average of 100,000s of tokens. The internet has 10s to 100s of billions of pages on it with who knows how many tokens on average. This is a huge dataset that we've got to go through hundreds of times.
Second, let's consider the effect of batching and how it sets requirements for our hardware. We know that larger batch sizes converge faster, are more stable, and produce better models. So if you want a good model you need large batch sizes. This means that you need machines several orders of magnitude more powerful than you use for inference. From what I heard Google uses clusters of 100s of the their TPUs all located in a single rack for training. These clusters are organized in a customized computing architecture to maximize memory locality between cores (really critical for efficient back-propagation). Further, you can't use reduced precision weights for training like you can for inference, so there are no shortcuts.
Finally, the initial training stage is followed by reinforcement learning stages - this is key development in how AI models have improved in the past year. This may mean going through a curated set of traces (either synthetic or captured from users) and adjusting the weights based on experienced outcome.
Overall there's so many orders of magnitude more work and more hardware requirements for training that I find it improbable that inference dominates. The number of "inference" steps in training is freaking ridiculous and includes such factors as the "number of words ever written".
It is easy for me to change providers. Right now I use the open source Claud Code harness with two paid API venders for DeepSeek v4 (flash and Pro). I like seeing how much each session costs.