i've been in this community a long time, mostly lurking and sometimes sharing what i'm working on. today i wanted to share something bigger. about a year ago i started on one question: can you train a neural network whose weights are just -1, 0 and +1, without all the giant memory that normal training needs? i didn't have money for real GPUs, so i bought a pile of old Nvidia K80s (Thanks @Aaron Quiroz) and built my own test rig out of them. that "can't afford the hardware" constraint ended up shaping everything. what came out of it: - SR4, my own training optimizer. its entire training state per weight is one 4-bit number, about 24x less optimizer memory than Adam. it matched full-precision accuracy on MNIST, and on real language models i've cut its extra cost vs Adam from 16.7x down to 4.1x so far - Pantheon + Foundry, my own inference engine and ML library, written from scratch in C with my own driver for the cards - a 35B MoE running on a ten-year-old K80 at 30+ tok/s in testing (this will ship soon for those following) then the work caught the eye of Tenstorrent. they sent me hardware in exchange for working on their stack, so i brought my engine up on their Blackhole card with my own driver and no tt-metal underneath. that led to: - a 9B model (Ornith) running on a single Blackhole card at ~50 tok/s, at 8-bit with exact output, no heavy quantization tricks - proof that ternary models can be stored bit-exactly in the Blackhole's native 2-bit format. a ternary model would fit about 100B parameters on one $1,400 card without losing anything i open sourced that work, and that's where things changed. here's why this matters. training big models is expensive mostly because of memory: with Adam, a 27B model's optimizer state alone is ~320 GB, which is why it takes racks of $25,000+ H100s. with SR4 that same state is ~14 GB. put that together with a $1,400 Tenstorrent card that runs ternary models natively, and the path opens up to training frontier-sized models on budget hardware instead of a datacenter. it's not free: SR4 still needs more training steps than Adam, and closing that gap is the core of the research. but that's the bet.

I just saw a YouTube video that discussed the death of MD Files, arguing that HTML files are the better way to go. Wanted to get the community ‘s thoughts on this and how it would impact ICM approach, if at all. Thanks

I will make a bet. Every knowledge worker will soon all be using a VSCode like app. Because otherwise, they will be replaced by one who does. Imagine every worker connected to the company data, controlling an agent that can also act on his desktop. The working app becomes: A. Files displayed in the middle of the screen B. Chat window (either left or right of the screen) C. Folder directory (either left or right of the screen) That's just VSCode. Not saying it will be literally VSCode used, but a similar app for sure. After some discussions with some very smart people, I was convinced that VSCode is the desktop app of the future.

We saw this coming for months.. I thought something felt off the last couple of days... Anthropic has officially renamed its $200/month Claude Max 20x plan to Claude Ultra. When Anthropic launched Max in April 2025, it advertised the $200 tier as offering up to 20x more usage than Pro. The new Ultra plan is now officially listed as providing 10x Pro’s weekly usage. Anthropic’s email says your “weekly usage limit stays the same.” That may mean the reduction had already been rolling out—or that the practical weekly cap was changed before the new name became official. Either way, the customer-facing offer has moved from **20x to 10x**. And we’ve had signs this was coming: - Anthropic introduced weekly limits in August 2025 - Users have increasingly reported hitting the Max 20x weekly cap much sooner than expected - Recent independent testing indicated Anthropic was A/B testing different Max 20x limits across accounts - Some users reported their weekly capacity already felt much closer to 10x than 20x Now the new plan documentation makes it explicit. What changed - Max 20x is now called Ultra - The price remains $200 per month - Ultra officially includes 10x Pro’s weekly usage - The rolling 5-hour limit is being removed for verified accounts - Usage can be spread across the week instead of being constrained by 5-hour windows - Max 5x is now simply called **Max** and remains $100 per month Removing the 5-hour limit is genuinely useful. Long Claude Code, Cowork, and background tasks should be less likely to stop halfway through a session. But we shouldn’t confuse greater flexibility with greater capacity. Anthropic is removing one restriction while formalizing a substantially lower usage multiplier than the one attached to the original $200 plan. The company can call it Ultra, but the underlying offer is now 10x usage for the same price previously associated with 20x. My read: the name change lets Anthropic retire the “20x” promise while presenting the transition primarily as an upgrade and silently lowering usage limits.

How do you deal with the AI overestimating the time/price a new type of project you're not familiar with already? You cut it by X amount? How do you approach it? Thanks