m_ke

Anthropic should think hard about all their fear mongering. It will only end up backfiring on them and everyone else involved.

They definitely used closed private saas products to train their own models, to prove that just drop random small screenshots of any popular product behind a login screen and see how well it's able to identify all of them. ex: https://x.com/michalwols/status/2079968211865330165

or other similar "AI" startups https://x.com/envconfig/status/2079613455296827402

I assumed google would lean into the efficiency stuff more and try to eat the easy 80% of workloads, winning market share on volume instead of frontier if they were not able to produce frontier level models.

They're very well equipped to be the volume discount store of inference.

k3 costs will go down at least 3x within a week of the weights dropping.

we'll get new quants, dspark speculators, distills and optimized kernels

as long as there are near frontier models available there will be inference providers selling them at or below cost of inference in attempt to get market share.

Sure, but anthropic is charging businesses based on usage now and tried hard to pull Fable from the consumer subscriptions before Sol and K3 dropped.

Even now on the $200 plan I use up my Fable credits in a single day and had to start using codex and openrouter for more usage because Fable burns $100s an hour when billed on usage.

Anthropic will get squeezed by open models for 80% of the use cases that don't require frontier capabilities and by vertical specific labs for the high value tasks that would (bio, finance, math, etc.), where smaller use case specific models will beat them on cost and speed while matching or exceeding the performance of their largest general models.

Even their hail mary of being first to "AGI" will never happen because all it takes is China blockading Taiwan or Nvidia cutting them off to stop them from eating up a large chunk of the economy.

There is no scenario where the rest of the world will sit on their toes and let OpenAI or Anthropic monopolize "AI". Too many countries, large well capitalized players and partners / suppliers who could never let that happen.

Kimi K3 allows all existing players to restart at the frontier and keep competing with OpenAI/Ant. It also gives employees at these labs a better more lucrative path of starting new labs with fresh books and clean cap tables, building on top of K3 without needing to spend all the capex on pretraining their own models. Plenty of them already vested their stock and would have 0 problems raising 100s of millions of dollars for new labs, making them paper billionaires over night.

most of the gains from the past year and a half have not been from web data, but from synthetic data and agent rollouts with RL.

yeah I tried rules, hooks and forbidding things like pip / python3 but it just led to the model failing to do what it wants and a bunch of token churn trying to get around my more rigid constraints.

main problem is that the harness files get loaded into context early in the session and slowly wash away as new information comes in.

It's not an issue of usual vs unusual, I'm saying the models are way better at writing and maintaining django or react code bases than your own hand rolled architecture that you define in some docs that it has to learn and keep in context. All of the models do an amazing job making local edits and working in small greenfield projects but once you get to full production systems with close to a million lines things start to rot. The code still works and the models are able to make progress but all of a sudden you have 3-4 different versions of your concepts sprinkled in random corners of your code base.

My second point is that the models are way better at things like Rust or Lua than Python or JS, because the average person producing code in those languages has way more programming experience, so the code quality of training data online in those languages is higher than the programming 101 medium blogspam type content you see in more popular intro languages.

I tried really hard to do this, but it turns out the models don't care about your ideas and want to do what's popular in their training data, so they will happily ignore anything you try to force down their throats, especially as context length grows or if you hit compaction.

So to make best use of the models steer them down familiar paths, mention common pattern and frameworks, use popular packages and languages that have the high median quality online.

I started my project with a few simple interface definitions and a short design / architecture doc that I include in the AGENTS.md file, but no matter how hard I try all of the models just end up ignoring it and sprinkled new seeds of variants of the same stuff all over my code base, that with each new session grow new branches.

With agentic coding the costs of tokens compound with each message / tool call and etc. Having to load in and update large files makes things slower and way more expensive.

Databricks actually just posted some of their own benchmarks on how harness alone impacts costs https://www.databricks.com/blog/benchmarking-coding-agents-d...

simple things like passing more file context, model having to explore the code base at start of each session, writing comments or markdown docs ends up increasing, running into test / build issues can 3-10x your costs.

PS: my code is still mostly TS and rust but I'm considering moving some of my annotations into .d.ts files and having them generated from runtime types (ala MonkeyType).

After running out of Fable credits in a day on my max plan I started looking around for ways to trim down my token usage and came to the realization that all of the type spaghetti that opus wrote is probably eating up like 50-70% of my tokens.

A clean django project is probably 3-4x less code than the equivalent TS based service.

It made me consider dropping strict mode and defaulting to js for most simple things.

Deep Learning models are designed to get max throughput on GPUs, which ends up being batched workloads.

You'll never get proper price competitive utilization on personal hardware vs a cloud inference provider that can batch and pipeline requests optimally to maximize utilization, unless you yourself start running batch jobs.

Even once local hardware and models catch up to todays frontiers, by that time there will be 10x better cluster sized models available at a similar discount.

Model is the product, people will put up with anything as long as you're on the pareto frontier of performance x cost.

https://openrouter.ai/models

you can use these in hermes, cursor, openclaw, opencode, etc with 2 lines of config that claude code will happily do for you if you ask

GLM 5.2, deepseek 4 Flash and the newly released Hy3 are Opus 4.8 and Sonnet 4.6 level models at a tiny fraction of the cost.

I'm on the $200 claude plan and blew through my weekly limits with Fable in a day, then ended up wasting $20 with opus 4.8 overages in an hour to finish out work in active sessions. Since then I've been using GLM 5.2 with openrouter + opencode and am spending less than $5/day for equivalent output.

It's looking like Anthropic is realizing that they're about to get squeezed so they need to juice revenue for their IPO before it becomes obvious to everyone else.

Open models like GLM 5.2 are getting good enough to handle 90% of tasks, and will eat most of their usage unless they start serving it at cost. And on the 1% work they fearmongered their way into falling under government control, which will limit how much they can commercialize the frontier.

Nobody will keep paying their premiums and put up with their BS when they can have similar models at cost of inference in any harness that they want.

I mostly use claude code on mobile and desktop with remote sessions and do most of my coding on the go now.

I have been tempted to distill something like GLM 5.2 into a smaller html + css only model for super fast interactive UI editing, because right now it's really annoying to do with large slow models. I'm sure it would be doable to do the same for individual language / frameworks, including potentially doing a final few steps on your own code base with some LoRAs that could be kept up to date to avoid having the model have to explore the code base each time.

Doing UI work with composer 2.5 and live reload is a way different experience than slogging through it with opus 4.8

i love when these errors bust my long running sessions and render them unusable

until it doesn't.

scaling laws are a power law, you can only stay ahead for so long when each minor improvement gets exponentially more expensive

see elevenlabs as a prime example

more like 0, once they get AGI they'll capture all value themselves or sell to highest bidders

So just a typical dropshipper

just think of what could someone do if they got into your personal email account?

No these fascists will celebrate this as not being woke, this is just them defending "the west" from muslims...

Guess who is running product and other related functions at OpenAI and Anthropic now

having been an early employee and founder of a few startups and then working at a few larger companies, most people who only ever worked at FAANG have no idea how much more productive tiny teams with ownership are.

I think the DNC and the media might need to get some of that blame for being empty vessels for corporate interests that allowed this conman to get elected twice

Don't blame the voters, they didn't get to pick her and did not run her campaign.

We can all thank the VCs and CEOs who fully embraced and enabled this administration

Imagine if OpenAI fails one day and sells to a company like Palantir. What would happen to all of the sensitive data that they're sitting on?

It's more than that, supposedly Sama donated another 25mil through a PAC.

I'm sure the Crypto AI Czar (David Sacks) being a major Anthropic hater didn't hurt either

Or that Kushner put a billion in OpenAI recently

EDIT: wow they got in at a huge discount too and OpenAI bought stake in Thrive...

https://www.wsj.com/articles/thrive-capital-bought-shares-in...

https://openai.com/index/thrive-holdings/