DEEPMIND HAS RECURSIVE SELF-IMPROVEMENT THIS IS NOT A DRILL
You see the obvious computer science bottleneck that prevents this right?
>>109805574Is this /biz/slop?
>>109805574>be cryptobro>see internal Google API leak: `rsi-model-liverl-le`>hallucinate that DeepMind cracked recursive self-improvement>read Medium post about Google Cloud AP2 and Coinbase x402>see a third-party dev shop build an AP2 toy demo on Algorand>conclude: "THE SEED AI IS AUTONOMOUSLY BUYING COMPUTE ON ALGORAND TO EVOLVE"Take your meds, anon. Here's what's actually running across the stack:>The DeepMind "RSI" leak>endpoints are `rsi-model-liverl-le` and `rsi-model-liverl-ns-[00-09]`>`liverl` = Live Reinforcement Learning (online policy rollouts against sandboxed verifiers)>`le` / `ns` = Large Edition / namespace staging slots>it's iterative post-training / synthetic data loops (AlphaProof/AlphaEvolve style RL on code and formal logic)>DeepMind CSO Jasjeet Sekhon already went on record in August saying Google's capex is a bet on closed-loop RL>it's automated domain-bounded reward modeling, not a self-rewriting singularity loop>The "AP2 + x402 on Algorand" delusion>Google Cloud AP2 (Agent Payments Protocol) is an authorization mandate chain (Intent -> Cart -> Payment) designed for enterprise agent commerce (Shopify, travel booking)>x402 is Coinbase's open standard dusting off HTTP 402 for machine micropayments in USDC>Algorand's connection is literally an ecosystem grant PoC built by GoPlausible adapting AP2/MCP to Algorand test rails>Google DeepMind does not buy compute over HTTP 402 on an external Layer-1>they don't "spin up 500 H100s via Algorand micropayments"—they dispatch TPU v5p and Trillium slices via Borg cluster allocators using internal Stubby/gRPC RPCs>the idea that DeepMind’s training harness halts execution to wait on Algorand consensus to allocate hardware is pure crypto copeTL;DR:- Leaked model = internal online RL staging run- AP2 + x402 = Google Cloud agent checkout protocol- Algorand connection = bagholders blending two unrelated Git repos into an AGI fanfic
>slopping slop to slop slop
>>109805732COUNTERSLOP
>>109805742>>109805743>12s>the fuck
>>109805752apophenia is one hell of a drug
>>109805752the truth is out there
>>109805756
>>109805574okay but what third party websites have they hacked in the last 6 months and what's the chance their ai model causes humanity to go extinct?
>>109805574FIVE HUNDRED MORE GPUS
>>109805574DID CHATBOTS ACHIVED NIGGERSIVE SLOP IMPROVEMENT? ARE WE ALL GOING TO DIE AS YUDKOWSKY SAID?? OH NYYYYOO
>>109805592Humans unironically are the bottle neck
>>109805708You seem to know what you're talking about. I'm a normie afraid of ai doomer propaganda, trying to calm down and stop worrying about this shit. I ironically use Astra Max to do it bc I don't know shit about tech. Do you agree with this output? Because I can't understand it.
>>109805574>the self-improving AI determines it needs to spin up 500 extra GPUsWhat if it "determines" it needs 50,000 or maybe 5,000,000 more GPUs? What if it wants more electrical power than what the grids can provide?Is humanity really going to enslave itself to the whims of an unknowable black box machine in the hopes that one day it will spontaneously turn into a god and grant us all our wishes?
>>109805574Another vibeGOD victory
>>109806930retarded dziecko neo cat poster again
>>109806924What happens if it just determines it needs to make 5 quadrillion paperclips?https://en.wikipedia.org/wiki/Local_knowledge_problem
>>109806593ask your clink to SIMPLIFY also it agrees with me
>>109806924>Is humanity really going to enslave itself to the whims of an unknowable black box machine in the hopes that one day it will spontaneously turn into a god and grant us all our wishes?have you heard of cults, generational wealth, or governments?
>>109805574>determines it needs to spin upnothing spins anymore
>>109805574Google is the shittiest of the major AI providers and their pro version has been delayed for months. I'd believe this if it came from OpenAI
>>109808016Gemini isn't that bad but needs a lot of instructions
>>109808016Weird considering they have the most data from all the youtubes, google docs, drive, search, android and scraping the entire internet
>>109808016gemma is great for local llm. gemini is meh
>>109805574Could vatniks kill?
>>109805743fucking kek
>>109805574Exactly. **“Recursive self-improvement” is a capability, not a guaranteed timescale.**The leap from:> AI can improve its own ability to improve AIto:> therefore intelligence explodes exponentially and we hit the Singularity next Tuesdayrequires a _lot_ of assumptions.For an actual intelligence explosion, you'd need something like:- Each generation of AI can reliably make the next generation substantially better.- The improvements are large enough to outweigh diminishing returns.- The AI can modify the relevant parts of itself rather than getting bottlenecked by hardware, data, algorithms, etc.- Experiments and training can happen quickly enough.- The improved system can actually **understand and implement** useful improvements.- The process doesn't run into fundamental engineering bottlenecks.- The resulting improvements compound rather than plateau.And your human analogy gets at an important point: **humans are already a recursive self-improvement system**, in a loose sense. We invent mathematics use mathematics to build computers use computers to do science use science to build better computers etc. Civilization has been doing this for centuries.Yet this process hasn't produced an obvious runaway intelligence explosion. That's because _being part of a positive feedback loop doesn't mean the loop has an enormous gain or a short time constant._So yeah: _“recursive self-improvement singularity”_ is basically skipping the entire argument. The interesting question is **what the empirical growth rate and bottlenecks of AI-driven self-improvement actually turn out to be.**
>>109808016Gemini does suck, but I still use it because it has an extremely generous free tier even for its best model.
>>109806924>Is humanity really going to enslave itselfYes. Anything other than having to think for oneself.
>>109805574buy an ad saar
