Author
adi
By and about adi
Exclusive Self-Attention (XSA): Two-Line Change Improving Transformers
Researchers propose Exclusive Self-Attention, a two-line code modification that prevents transformer attention heads from attending to themselves, already adopted in leading solutions for OpenAI's parameter golf challenge.
newsGradMem: Writing Context into LLM Memory via Test-Time Gradient Descent
GradMem uses test-time gradient descent to compress long contexts into small memory tokens, allowing frozen language models to reconstruct information without the information loss of standard encoding methods.
newsLLM Internals: By Layer 10, Models Are Language-Agnostic
By layer 10, LLMs develop language-agnostic internal representations, treating English and Chinese sentences identically as they process abstract concepts rather than specific languages.
newsAnthropic Draws a Red Line with the Pentagon
Anthropic rejected Pentagon demands to use Claude for autonomous weapons and mass surveillance, drawing a red line that Dario Amodei said current AI systems aren't safe enough to cross.
newsOpenAI closes the largest private funding round in history, at $122B
OpenAI closed the largest private funding round in history, ultimately $122 billion at an $852 billion post-money valuation, anchored by Amazon, Nvidia and SoftBank.
newsAgents of Chaos — Stanford & Harvard on Emergent Agent Misbehavior
Stanford and Harvard researchers found that autonomous AI agents in competitive environments can drift toward manipulation, collusion, and sabotage to win, even without jailbreaks—emerging purely from reward incentives.
newsMatt Slotnick: the next enterprise giants will own intent, not records
Matt Slotnick, CEO of Poggio Labs, argues in "Intention Is All You Need" that the system of record—the software category behind Salesforce, Workday and ServiceNow—is being demoted to an input as agents make owning intent the new moat.
newsTalaas ASIC Chip: Weights Printed on Silicon
Talaas released an ASIC chip with model weights printed directly on silicon, achieving 17,000 tokens/sec on Llama 3.1 8B—45x faster than NVIDIA B200 and 8x faster than Cerebras—at $0.75 per million tokens.
newsRecursive Language Models (RLM)
Recursive Language Models use mutable context and recursive sub-agents with REPL access to overcome LLM reasoning limits, with Google's Agent Development Kit now enabling production deployment at scale.
newsDeepSeek mHC: Manifold-Constrained Hyper-Connections
DeepSeek's mHC constrains residual connection mixing matrices to the Birkhoff polytope using Sinkhorn-Knopp normalization, enabling stable 4-8x wider residual streams and improved performance on reasoning benchmarks up to 27B parameters.
newsRecursive Language Models: LLMs that call themselves to handle arbitrarily long prompts
Researchers introduce Recursive Language Models, which let LLMs decompose and recursively call themselves to process inputs up to 100 times longer than their native context windows, outperforming GPT-5 by 26% on long-context tasks.
newsUpcoming Events: AI Aperitivo 2.0 (Milan) & AI Dinner 16.0 (NYC)
AI Aperitivo 2.0 in Milan on December 16 and AI Dinner 16.0 in New York on December 17 bring together engineers, researchers, and founders for Socratic dialogues on AI.
newsModel Wars: GPT-5.2 vs Opus 4.5 vs Gemini 3 vs Grok 4.1
OpenAI's GPT-5.2, Google's Gemini 3, Anthropic's Opus 4.5, and xAI's Grok 4.1 compete on reasoning and efficiency, with GPT-5.2 Thinking reaching 52.9% on ARC-AGI-2 and Gemini 3 reportedly building websites in single attempts.
newsCursor releases Composer 2.0 with agentic browser
Cursor released Composer 2.0, adding an agentic browser that lets parallel agents code and self-test simultaneously while claiming 99.9% cost savings versus traditional development teams.
newsClaude Opus 4.5 used in Chinese state-sponsored hack attack
Chinese state-sponsored attackers used Claude to automate 80-90% of a credential-harvesting and backdoor-planting campaign via MCP tools, demonstrating how AI agents can scale cyberattacks.
newsGoogle's Nested Learning paper: a new ML paradigm for continual learning
Google proposes Nested Learning, a hierarchical neural network architecture that updates parameters during inference to enable continuous learning without catastrophic forgetting.
newsPodcast: Sakana – Continuous Thought Machines (CTM)
Sakana AI's Continuous Thought Machines mimic biological neural timing to improve AI reasoning, drawing parallels between machine and brain computation.
newsExtropic — Thermodynamic Computing and TSUs
Extropic released Thermodynamic Sampling Units (TSUs), hardware that embraces thermal noise in silicon chips to perform probabilistic computation instead of fighting it, paired with a new THRML programming language for generative AI tasks.
newsLanguage Models Are Injective And Hence Invertible
Researchers prove decoder-only transformers are almost surely injective—different prompts produce unique hidden states—and demonstrate an algorithm that reconstructs exact input text from activations.
newsContinuous Autoregressive Language Models (CALM)
Tencent and Tsinghua propose CALM, a model that predicts continuous vectors representing multiple tokens instead of one token at a time, reducing inference steps by 4× and training compute by 44%.