If you continue to do this and adopt more compute, you might just become the biggest token plan provider for open models..
-----
It would also help if you guys launch a idle time product like dots/code review/security testing or just scheduled jobs, but it gets complicated with managing VMs and all that shit, but might be worth it if you got all those cores sit idle from the GPU boxes..
@Elyte7894@opencode With the worst cache hit I ever seen. Leading to me consuming 30$ for something I pay Ds v4.1 flash thru opencode at about 0.25$ max
@wcblessing@Im_pritam18@grok I had best experiences with perplexity when they started and that was mostly RAG back then. They never told me something that was incorrect factually or a hallucination. I wouldn't call that just hype..
Are you ever going to maintain/improve that codebase without AI?
I think having a comprehension of the following is more important:
What are the entry points into this service?
Do the entry points do what we want?
What is a threat model and what is a customer model?
Do we have sane authentication over everything considering those two models?
What is org compromise? How many steps would our org take to get there? Can we implement something to make isolation of security incidents to lowest amount of data possible, maybe even no data? and an alert to be triggered when it happens?
And the nice part, all this can now be done properly by a very small team.
So transformer models are essentially turing machines, your input is also part of the code and because of that all models can actually be steered, now imagine someone finding a common pattern of words that make a behaviour more proeminent or less proeminent. Now add to the prompt the self replication part that it needs to spread and you are done :)
It's crazy that we have not found a self replicationg prompt injection yet that is widelly working across model families.
https://t.co/tvitxK7E7A
It's not even about that somehow, but more about how one poisoned model can try to proactively poison other models thru his action and what is the spread of this and how it can be stopped.
Cyberweapon level ai models are already here..
Now imagine a army of them and also the misalignment that a mind virus can cause.
Also, mind viruses can be simple, they may look like normal words, if you study this a bit you will understand my point.
@kannthu Can we share some ideas on this one? I done something very cool with esphome + wifi for p2p and I don't really have the time to take it up to next level. Maybe something comes out, DM me if interested
@kannthu You can just tell them the ai escaped out of the sandbox and that you are sorry it sol all their customers data on the dark web to make money to train better models .
I guess it comes with AI autonomy.
@weezerOSINT Mistral is shit tbh, has low cache hit and costs a bunch, opencode go is the way for DeepSeek. Glm 5.3 costs too much also...
For finding bugs, DeepSeek still rules at 1/30 of glm 5.3 cost ..
Compaction is a construct given by the attention mechanics of llms and for that is the most efficient construct.
Forgetting is continuous and we do it by moving some stuff in persistent memory (skills and such). And at night we run benchmarks for what we learned to make sure it's not misaligned so we can push to prod.
That's why if learning was not successfully.that day you might need another training round.
Whoever escapes attention can get rid of compaction cheaply, but that does not seem in view yet on research side sadly.
Maybe someone could hire me to solve it.
Cc: big ai companies hahaha