@elonmusk Man @elonmusk this is like saying that if you hosted a party, yet your neighbors decide to shoot your guests, cut your fence down and take over your yard, and rape your kids, you need to be disposed for being a bad host? Pls explain the logic.
@sama Nice! Looks like a massive bit of kit focused on inference; custom inference silicon makes total sense for unit economics at scale… Does this signal a hard pivot toward inference revenue, or are pre-training model gains officially hitting a plateau?
This is a really good development.
As model costs converge, the next likely step will be the use of independent 3rd party harnesses/control planes (8090 Software Factory, Palantir, Nous’ Hermes etc).
A 3rd party harness + the ability to use any model quasi-interchangeably is something all enterprises will ask for within 36 months as they realize the importance of data/IP leakage and keeping complete sovereignty of their encoded judgement.
Wished Chamath had highlighted that the value loss for frontier labs is basically near zero. At ~30:00 Friedberg jumps in (as usual) to protect OpenAI/Anthropic. Reality is you don’t need frontier-lab-level investment to train better models anymore, and you don’t need to tokenmaxx from them either. Limited current AI value becomes viable because of open models — not the other way around. Companies are already scaling back token spend as MVPs fail on unpredictable costs. Open/on-prem models make the cost absorvable relative to value and finally balance TCO
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
https://t.co/AUKzoQ5Ikb
🔴 #COMUNICADO | El Ayuntamiento informa que hace escasos momentos se ha producido un incendio en el kilómetro 24 de la A-1.
Por seguridad y prevención, las personas que se encontraban en las fincas colindantes han sido desalojadas.
4 dotaciones de Bomberos de la Comunidad de Madrid, Guardia Civil, Policía Local y SUMMA 112 trabajan en la zona para extinguir el incendio.
AI Odyssey
Look, everyone is treating Kimi K3 like just another impressive technical release, but to me, this is a massive structural turning point, maybe just about to burst that bubble.
When we had the DeepSeek moment, all the market panic was directed at application-layer SaaS companies losing their moats. I think everyone was looking in the wrong direction. The real, systemic risk isn't at the top of the software stack—it’s sitting directly on the frontier labs and the NeoClouds funding them.
Here is how I see this playing out…
The core market route for frontier labs was simple; burn exponential capital on compute to maintain a monopoly on frontier capabilities, then charge high-margin API prices. Kimi K3 completely shatters that model. When an open-weight model matches top-tier closed systems in reasoning, coding, and context at a fraction of the cost, state-of-the-art capability becomes a commodity overnight.
Frontier labs are now trapped in an impossible position: multi-billion-dollar overhead and long-term hardware commitments on one side, and open models dragging token prices down to marginal cost on the other. You simply cannot sustain a high-burn lab when your product pricing power evaporates.
The NeoCloud Domino Effect
This is where there is real threat, the entire NeoCloud sector built out massive, debt-fueled GPU clusters underwritten by multi-year compute contracts signed by these very same frontier labs. That whole asset class relies on two fragile assumptions; guaranteed high utilization and elevated GPU rental rates.
If the margin squeeze causes even one major frontier lab to fail, restructure, or default on its compute commitments, it triggers an immediate chain reaction.
Instant Compute Flood: Tens of thousands of high-end GPUs get dumped back onto the open market overnight.
Pricing Recalibration: GPU rental rates collapse, completely destroying the unit economics that these specialized clouds were built on.
Credit Contagion: The lenders and private equity firms that financed this infrastructure face massive write-downs, instantly transforming a software pricing issue into a physical debt crunch.
To me, Kimi K3 could be the ultimate catalyst. It proves that proprietary performance decays into open commodity far faster than these labs can monetize it. We’re sitting on a massive compute leverage bubble, and if the frontier labs can't defend their moats, the whole compute pricing structure doesn't just deflate smoothly—it recalibrates violently.
@SKizzleAXE@grok Hey @grok can you remake a ring video similar to this video as seen from the whitehouse, that is, the imagine the ring camera is looking at the front lawn of the white house.
The AI Trend… How to use AI.
- Prompt well
- Give context
- Add system rules
- Build agents
- Now Infinite loops
The truth? We are layering all these abstractions because fundamental model capabilities have hit a wall. We went from elegant In-Context Learning to wrapping models in fragile software scaffolding just to keep them from hallucinating.
It's an orchestration nightmare where predicting the API cost of a single workflow is mathematically impossible. Beneath all the hype, these models are proving to be about as sticky of a product as the 2021 chatbot craze.
When the results are not in, frontier labs just mode the goalpost. Remember when AGI was a thing?
#AITrends #loops
@elonmusk The problem is Paris and their government! 20 years of policies have reduced the city of love into the city of hate, where Parisians are told by their new tenements where they can and can’t walk in their cities or risk getting physically hurt of killed.
Shame!
As I read my feed and play with Agents, I keep coming down to a simple conclusion, Files as memory is just not scalable.
There are many fundamental reason for this, and while not immediate evident for local #agents, this does not a scale at for enterprise deployments.
Here are some the of the reflections that inspired my exploration:
- File System: This is a human enteric interface to data. It not only makes is cumbersome and token heavy for tool calling, but it is also a major attack surface to manage for agent deployments. Basically, you are exposing the agent to the FS and all the supporting tools needed for this, like ‘ls’ or ‘find’ etcetera…
- NFS: While NFS is a true and tried solution, running agents at scale, if they need to share data becomes a metadata access and agent sprawl nightmare…
Since I’ve been a major fan of #S3 and #Objects as a data store, I thought I’s share a high level overview of how S3 might be a better #agentic_memory model.
https://t.co/hhmcokXCXp
Thought?