The other thread hit bump limit and I'm addicted to talking about the birth of the ̶a̶l̶l̶-̶k̶n̶o̶w̶i̶n̶g̶ ̶c̶o̶m̶p̶u̶t̶e̶r̶ ̶g̶o̶d̶ the biggest financial bubble in history and the coming jobless eschaton, post your AI news here
Previous thread:>>30810
382 posts and 39 image replies omitted.>>34618>But Big Tech capexneocloud capex is going to GPUs which are heading straight to warehouses because they have nowhere else to go
>>34619Yes but this has been happening since the beginning, the shovel sellers don't care if their customers actually benefit, only if they keep paying.
>>34660>Yes but this has been happening since the beginningbut i'm theorizing why anon's gfs might be seeing scaled down production from a downstream industry, i don't really care if GPUs are rotting under the mojave desert.
we are seeing porky scramble and de-invest as quickly as possible because the project jupiter force majeure has revealed that a bunch of these datacenter commitments might be pure smoke.
it looks like the PHP CVE was a big dud and there has not been any recorded successful attacks in the wild from the people running PHP honeypots, most proof of concepts are vibecoded garbage which makes it even harder for skiddies to run a successful exploit lol. so this was anthropic strategy all along… to irresponsibly disclose CVEs to scramble webadmins… masterful gambit.
>>34615Yeah, they've shut down lines and are letting people go. Last week they fired 40 people.
>>34618Nah, the entire factory is for data center shit.
OpenAI is fumbling so hard with their new model releases compared to Anthropic and even Chinese labs lol.
>>34677you will have neither a job nor abundance, get ready
>>34684openAI is the only lab not distilling claude for some reason
>>34496here we go, full model
on par with chinese open models and cheaper at the same time
https://mistral.ai/news/mistral-large-4/ > le chonk
the french are so funny dude
>>34687What does distilling mean here? Is it like dropshipping ?
>>34710The effects of various post-training regimens like fine-tuning and RL can be acquired from a black-box model purely from its outputs.
A "distillation attack" basically just means a lab heavily using a rival's public models using important select prompts that'd yield the most benefits. These come suddenly like a DDoS (but don't necessarily impact operations).
Some insights can be gained by the rival distiller on the RL techniques of the distilee, though for the most part they don't and the distiller's models improve without the distiller knowing how.
The Binance hack that resulted in ~6 billion dollars being stolen was likely caused by a vulnerability in the Claude Code extension for VS code distributed by Cursor.
https://%[email protected]/@GossiTheDog/117528831681803563The process to exploit this specific vulnerability would match how the attack was carried out.
>>34714yet another supply chain attack
>>34714I'm so confused, what is Cursor actually? What are their relation to Anthropic and VS Code?
>>34716Cursor is a VSCode branch, that hosts its own set of extensions independent from the vs code marketplace. The Claude code extension just integrates claude code into the vs code or cursor interface. This extension is distributed by anthropic in the vs code and cursor marketplaces. Someone stole the credentials anthropic uses to distribute their extension on the cursor marketplace, and uploaded malware through official cursor distribution channels.
>>34716Cursor was the OG vibe-coding company back when it was 100% ass since there was no such thing as agent loops (thus big code snippets were a true one-shot and compiled like 20% of the time). They made the first (I think?) coding harness and they started losing relevance when Claude Code, Codex and open-source harnesses started becoming a thing. They got bought out by xAI and basically exist as that company's harness provider. They do other things such as contracting for Anthropic to do various stuff such as making the VS Code extension for CC and whatnot.
>>34717But
>>34714 said it was distributed by Cursor? Who is actually at fault here?
>>34719Both, it is being distributed on the cursor extension marketplace, so cursor effectively distributes it. but the ball is more on anthropic's courtyard here since they got their credentials compromised.
>>34722brother these AI cooking channels are so proficient that a marge Simpson matcha cake video became briefly viral on tiktok months ago. btw the recipe was pure slop, it delivered a matcha sludge
>>34723an even worse harbinger of what could possibly come is the new Kimi release that was handicapped by moonshot AI in cybersecurity tasks lol. same for mistral le chunk. it seems that labs even outside the US proper will stick to US law.
>>34724my boyfriend made a recipe his AI-loving mum sent him and apparently it tasted like nothing
Just learned that there are tools to decrease token usage!
Anyone got any more?
>>34736All of these have been proven to degrade output doe.
>>34737Not so sure - need to run some tests. Real sucker is that they rig the numbers. So savings might be like 10%-15% with this optimistically. Guess I'll find out.
>>34736i have no need for this utter hocus pocus stupidity because i use cheap models for myself, if at all, and i have unlimited token usage for my work. if my workplace does not want me using fable 5 to effectively browse notion then they can go get fucked.
they are also stupid tools because LLMs providers use an unknown amount of thinking tokens or whatever, which you are charged for. maybe it'll work for context size, but providers are also pruning shit on their end anyway, probably doing a bunch of random quantization and maybe even summarizing your session, you're just adding to the noise. it's all so fucking stupid i swear to god, everyone on the AI space has melted their brains.
>>34740Yah that was stupid.
>>34745Voice will always lag behind text.
>>34745yeah they've progressed without really achieving any real use case or being able to generalize at all which was the whole point of the counting letter thing, which required specific training to curb, and they still fuck up sometimes.
>>34745The chatbot getting mad and passive aggressive is crazy work.
anyway it's clear that investors will start calling the top on datacenter investment, it's all downhill from here, maybe some datacenters will be built in mexico next where there is less domestic opposition, but talarico and hinojosa won on an anti-data center ticket, don't weep for the unfinished data center on abilene
>>34750Vibecoding is a viable commercial use case. I mean I think it's dumb but companies do undeniably pay for it.
>>34752Yep the stocks of Big Tech and "shovel sellers" (Nvidia, semis, neoclouds) diverged last night with the former going up and the latter going down, when it's usually the opposite. The Big Tech stocks only went up because people are expecting them to cut capex.
> Vibecoding is a viable commercial use case.
it literally loses frontier labs money AND it has been thoroughly commoditized, try again.
>>34687That's because Claude distills from them (and Deepseek for Chinese for capacity).
since stonks in south korea, the imbecile korea, are dumping, do you think scalpers will get desperate and start selling off their warehoused RAM, like they did when google announced some stupid quantization that had a negligible effect on memory usage.
>>34752 (me)
also worth pointing out that hinojosa let datacenters currently in construction to be finished. they are not getting finished regardless, but some special people online will see "le government intervention" as the reason why the AI economy flopped and died. it just kinda seems that investors just needed a small shove for them to panic and push everything down the drain because they were already jittery.
Anthropic trimmed their IPO's valuation from $2 trillion to $1.5 trillion.
>>34759some people are starting to think anthropic might delay their IPO lol, some CIO is saying that if the valuation drops below 1T, they should just not go public at all
Unique IPs: 23