mirekrusin

because it doesn't suck :)

it's local first, agent-friendly tooling; you have single cli to access it (mcp is there as well if that's your thing)

everything is persisted in simple local text files that you can actually read (or sqlite database if you don't care / have more data)

ie there is no integration, k8s-style nonsense etc, you can just use it on project or on your system user level and know exactly what's there, you have control over it, you can delete it, move it, extract parts you find useful and move on etc

because it has first class support for uncertainty

evidence is first class concern

supports evolution of belief (statements can be revised, retracted etc. without destroying history)

knowledge is executable (rules, actions, automations – all operate on the same claim model)

it works well for knowledge that keeps changing (including support for querying knowledge as of some past time)

it's great for llm sessions, ie. simple things like comment on a claim being part of a claim itself, which means it appears when working with knowledge, which allows to support weird/human/natural-language aspects that are not directly part of the model

...

rdf/owl/sparql is great for large, public, structured datasets (ie. medical, scientific or some regulatory, standardized ontologies)

cave is great for evolving knowledge, ie. in most cases i use it by just dropping it in and forgetting about it – i don't care how/what llm will model, i just care it'll solve some problems faster when aiding itself with it as a tool; if i want some structured dataset out of it i may occasionally ask it to model it around homogeneous schema i can extract, but in most cases i really don't care

Yes, and few other useful concepts beyond just Datalog.

I still prefer my own, simple and more functional CAVE lang [0]

[0] https://mirekrusin.com/cave

...or ai lab bench maxxing.

They should do franken-MoE with JEPA head and LLM head.

No login/account? What a blasphemy. Feel in love from the first site.

To be fair it’s not like screwups were not happening before, but if you move 100x faster and are 50x more secure than before you’re still 2x more exposed.

Just to be clear – my comment was a joke.

Modding isn't equivalent to post-training. Modders write new code, replace assets, patch existing instructions etc. With model weights you can continue original training process – directly optimizing existing artifact to acquire new behavior. There's no equivalent general-purpose process for a compiled binaries.

Same as "airplanes are made to be used as submarines" – you just need hammer and few other tools and you're good to go.

You're conflating runtime patching with continuing development.

Post training directly evolves existing model weights using the same optimization process that created them.

Software: source code -> compiler -> binary

Models: weights + training data + optimizer -> updated weights

Ie. binary is output of the development process.

Model weights are both – it's output and it's direct input for further development.

Model weight snapshots are continuously used all the time in this iterative process from the very beginning, being output of current step and input to the next step.

Post training is general purpose method, same process that original authors used - that's not the case with binaries.

Do you have examples you can point to?

For example there are more than 100k modified weights on hugging face.

I think they were expecting them to be replanted on the roof.

Can you evolve binary like you can post train weights?

It's not them who decides if they can release it to the public but US government.

Once government clear it, it'll become available.

Same delay process will happen to new frontier models from other US labs.

As long as it's from US lab, it'll also have to go through US apparatus.

well, they don't burn it like anthropic on rent, they spend it on their own TPUs, right?

Out of all companies, with their own TPUs and native large scale computing... "habitat" they exist in – they probably know how to spend it well.

Why their shares tanked 10% then?

Open-weight models

Kimi K3.1 Pro: −9

DeepSeek V4.1 Flash 1017: −13

GLM-5.3: −13

MiMo-V2.6-Pro: −16

Proprietary models

GPT-6 Astra: +4

GPT-6.1 Sol: +3

GPT-6.1 Terra: +3

Claude Opus 5.5: 0

Claude Sonnet 5.5: −2

Claude Fable 5.1: −3

Gemini 4 Argon (high): −3

Muse Wonder Full: −6

Claude Haiku 5.5: −7

Muse Wonder: −9

Gemini 3.8 Flash (high): −10

Grok 4.7 (high): −11

Is there affiliation with OpenAI or coincidence?

success and failure of anthropic are both good for open weight models

It goes down because contracts expire.

AMD is missing, true but commitment schedule isn't known. Also missing estimated additional TPU purchase for 28-31, ~$125B-$150B.

- 2026: $97.08B = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B

- 2027: $97.08B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + AMD (TBD)

- 2028: $128.33–134.58B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2029: $128.33–134.58B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2030: $98.21–104.46B + AMD (TBD) = Broadcom $26.87B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2031: $98.21–104.46B + AMD (TBD) = Broadcom $26.87B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2032: $40.09B = Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B

- 2033: $40.09B = Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B

- 2034: $14.59B = Amazon $10B + Lambda $4.59B

- 2035: $10.00B = Amazon $10B

- 2036: $10.00B = Amazon $10B

Combined total: $782–807B+, including estimated additional TPU purchases and conditional AMD spending.

1. it comes from VCs philanthropy, who cares?

2. it wouldn't happen so who cares?

3. no, but models will be better, it'll be cheaper to do

It's not a snark. The scale of VC philanthropy is unprecedented and it fuels itself. All of it will eventually, one way or another, become open. Long story short – there is nothing to complain about.

from what you've written it's more like:

- 2026: $97.08B

- 2027: $97.08B

- 2028: $97.08B

- 2029: $97.08B

- 2030: $66.96B

- 2031: $66.96B

- 2032: $40.09B

- 2033: $40.09B

- 2034: $14.59B

- 2035: $10.00B

- 2036: $10.00B

Total: $637.0B

Considering everything that is cancellable, ie. hard commitment only, it's:

- 2026: $56.79B

- 2027: $56.79B

- 2028: $56.79B

- 2029: $56.79B

- 2030: $56.79B

- 2031: $56.79B

- 2032: $29.93B

- 2033: $29.93B

- 2034: $10.00B

- 2035: $10.00B

- 2036: $10.00B

Total: $431B

If they go bankrupt, there are some more cancellations available, also from the other side – if some providers don't deliver, it can be lower as well.

I'm sure tools don't like annoying people either.

I'm not sure if not doing RL capability gain on offensive cybersecurity can be really called "nerfing"?

They're not post training it to lose capabilities/censor them – they just don't spend training time on this topic, which sounds like a great idea.

It doesn't mean model can't do it, it just means it's less good at it; -10% on benchmark feels like a great result and overall approach to pace it.

It would be great if numbers weren't off, would include optimised versions (ie. speculative decoding), context size, include allowed parallel inference (single agents use parallel inference a lot) etc.

I don't think in general people care about long list of models, they just care about the best one.

Also dual gpu is common setup, there should be ie. 2x 4090.

It's conducting/steering/oversight/supervision, not reviewing.

I think the consensus is to just keep calm and carry on, next model release will sort some of it out.