GPT-5.6 Sol, Terra, and Luna (39 minute read)
OpenAI introduced GPT-5.6 Preview, a family of models named Sol, Terra, and Luna, with Sol positioned as the flagship model. The system card describes stronger cyber and bio safety testing, new safeguards, and a limited preview before broader availability.
|
Musk Says Grok 4.5 Entered Private Beta (1 minute read)
Elon Musk said Grok 4.5 was in private beta at SpaceX and Tesla. The model is based on a 1.5T V9 foundation model with Cursor data added during supplemental training. Early evaluations were near or above Opus, with reinforcement learning still improving the model.
|
Google Limiting Meta's Gemini Use (2 minute read)
Google reportedly limited Meta's access to Gemini capacity after Meta requested more compute than Google could provide. The shortfall was said to have delayed some internal Meta AI projects and pushed staff to use AI tokens more efficiently.
|
|
Lean Software Scaling Laws (17 minute read)
Codebases and programming languages will eventually become easier for AI models to understand, fix, and write. The Lean programming language has a worse baseline constant and total loss on existing code bases compared to other languages, but better scaling components. This implies that implementations in Lean could eventually win and deliver large benefits in program correctness at global scale. This may justify large-scale investments in rewriting existing codebases in Lean or paying for new Lean code.
|
Memory Prices report from Stanford (4 minute read)
Stanford published an interactive report on historic and current memory and storage prices. The dataset follows the spirit of John C. McCallum's classic memory-price work, with downloadable raw data and chart export tools.
|
The Next Paradigm (7 minute read)
AI labs are betting that scaling reinforcement learning from verifiable rewards (RLVR) across millions of diverse tasks will achieve artificial general intelligence, but this paradigm hits a wall in domains that lack deterministic simulators. True continual learning requires moving past temporary in-context memory and shifting back into the model's weights themselves.
|
Moneyball for Physical AI (26 minute read)
Data engineering pipelines should deprecate cumulative operational hours as a primary metric. Engineering efficiency and model scaling should be evaluated using quantifiable parameters. An optimal capital allocation strategy balances data types against their specific utility metrics. Capital efficiency scales by accurately pricing data novelty.
|
|
Stuck in a botsitting cycle? (Sponsor)
The Work AI Index 2026 from Glean's Work AI Institute surveyed 6,000 digital workers and found that AI saves time, but much of it goes back into cleanup. Read the report to see where AI time savings go and what high AI achievers do differently.Download the report
|
Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction (10 minute read)
Models like Gemini Nano and Gemma make it possible to have powerful large language models right in your pocket. Delivering these models on mobile was a significant challenge. Google built a new architecture that retrofits Multi-Token Prediction onto existing 'frozen' Gemini Nano v3 models to overcome the bottleneck. The new architectural components were designed to maximize efficiency gains specifically for mobile environments. This article shows how Google's research team tackled the unique, extreme constraints of edge computing.
|
Qwen Image Agent (12 minute read)
Qwen-Image-Agent improved text-to-image generation by planning, reasoning, searching, using memory, and incorporating feedback to fill in missing user context. The work also introduced IA-Bench to test agentic image generation across planning, reasoning, search, and memory.
|
Reward Models Can Be Too Sensitive (22 minute read)
Meta studied how reward models can overreact to equally good responses, leading reinforcement learning toward reward hacking. The paper proposes measuring both discriminative ability and specificity, then using Monte Carlo dropout to cluster rewards into safer discrete signals.
|
|
Agents as Webs of Beliefs (11 minute read)
A proposed framework models intelligent agents as interconnected webs of beliefs, where beliefs, goals, and actions emerge from the same underlying structure instead of being treated separately. The approach argues that reasoning, planning, and decision-making arise from maintaining locally consistent belief networks, offering an alternative foundation for building more capable AI agents.
|
|
|
| |
