Super Intelligence.
s.i.
sī - 思 - to think, consider, or reflect
sǐ - 死 - to die, dead
... very ominous
Usually the ugly bits that actually warrant such opaque implementation can be quarantined beneath more legible abstractions
Even without a switch to latent space neuralese, I wouldn't be surprised if soon enough the majority of LLM thought happens in opaque on the fly invented and/or RL emergent information-dense languages
GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.
In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels. When we examined the reasoning chains to understand how the model operates, we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level. It goes as far as developing its own shorthand DSL to represent in-game situations -- essentially a game-specific algebraic notation.
Overall, Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses -- so harness capabilities are increasingly shifting into the model itself.
We see Astra as a major breakthrough in model intelligence.
Read our post on Astra and what these results mean: arcprize.org/blog/astra
Interesting how the LLMs freely coin new terms or new context-specific meanings for terms as an information compression mechanism
I wonder what is the Shannon capacity of the Claude watermark -> distillation model output channel
Anthropic says new Claude models will embed invisible watermarks in all generated text, everywhere Claude is offered.
The watermark is part of the text, it isn't metadata: "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."
This starts with models launched on or after August 2, 2026, under an EU AI Act code Anthropic signed. Anthropic is still working on adding it to current models. The rollout is worldwide.
Dario is signaling that he doesn't care what any of you think. And that's what you should expect from his company.
This quoted post is unavailable.
The knowledge in most domains is not derivable by pure computation/reason. Those will be limited by the bandwidth of channels built for acting on and sensing the physical world. Massive opportunities for teams closing that loop.
I find it extremely unlikely that exporting intelligence will be the main form of monetization for Frontier AI companies in the long run. More likely, companies will just in-house the actually economically valuable activities. Intelligence is just a means to an end.
These are basically factor graphs and this is the Forney-style representation.
You'll see similar in viterbi, kalman filters, modern error correction codes (eg LDPC), probabilistic graphical models, etc. Many of the algorithms in those domains are reducible to the same type of message passing within the respective graph.
Index sums vanish once tensors become circles with legs.
Scalars sit alone, vectors carry one leg, matrices two, higher tensors more. Group indices and the leg thickens. Join two legs and that index is summed away. A full network contracts into one tensor whose open legs are the free indices.
Matrix product states for quantum spin chains are contracted this way to extract ground-state energies on ordinary computers.
Seems to me that agents should be constantly redlining at nearly full context for best task performance, purging only the least useful information as necessary. What am I missing here?
