2/5 We started seeing real gains when we could take advantage of a diverse portfolio of models – e.g. the operating points for individual ensembles (running in batches & stopping when a candidate passes a self-confidence threshold) appear above the best single-variant Pareto
LLMS
-
AI21 Maestro Achieves SOTA on BrowseComp-Plus with 95.18% Accuracy
By
–
1/5 We hit SOTA performance on BrowseComp-Plus with 95.18% accuracy using AI21 Maestro’s agent optimization. Here’s how we automated the search space and reached #1.
-
Estimating Black-Box LLM Size via Factual Knowledge Probes
By
–
"Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity" This paper estimate close source LLM size from long-tail facts by using 1400 probes of obscure knowledge. The idea is that reasoning can be compressed, but factual storage can't.
-
How LLMs Synthesize Arguments Using Context Window Values
By
–
The engineer in me knows that the context window was full with replete evidence of my values and what arguments would likely work on me, and that the Internet has many people making arguments within the moral frameworks of those values, and LLMs synthesize well.
-
LangSmith Agent Deployment with Multi-Provider Model Support
By
–
[agent] This is where you specify the name of the agent – this is the deployment name in LangSmith This is also where you specify the model name. We support a lot of models! Not only ones from OpenAI, Anthropic, Google, but also @OpenRouter @FireworksAI_HQ @baseten @nvidia and
-
AI Demonstrates Clear Moral Reasoning in Unexpected Refusal
By
–
Today is April 30th, 2026, and today was the first day I can recall being impressed with moral reasoning articulated by a software product in refusal to do a thing I asked it to do. It misparsed my intent, and I do not *agree* with its moral reasoning, but clear and cogent.
-
Hassabis: AGI Is Close, Missing Continual Learning and Memory
By
–
Demis Hassabis: We're on the right track to AGI; we probably have all the components. We're just missing a few things like continual learning and solving the memory problem. pic.twitter.com/zuIKOiKnB7
— Chubby♨️ (@kimmonismus) 30 avril 2026Demis Hassabis: We're on the right track to AGI; we probably have all the components. We're just missing a few things like continual learning and solving the memory problem.
-
Preconditioned DeltaNet Adds Curvature-Aware Linear Recurrence Modeling
By
–
“Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences” This paper views linear recurrences through a least-squares/test-time regression lens, and adds the missing curvature information via preconditioning. Main idea: precondition the delta-rule
-
AI Assistants Still Corrupt Documents Despite Trillion-Dollar Scaling
By
–
Ouch! Current AI assistants often corrupt documents. Sounds like an intern you can’t trust — once again. A trillion dollar investment in scaling hasn’t solved this.
-
Small Models Handle Repetitive Agent Task Layers Efficiently
By
–
Any repetitive agent task layer – routing, validation, retrieval – small models can handle it