Notes on ai

Showing notes about ai

Topics

21 notes on this topic

ai (21)agents (14)models (9)memory (7)architecture (6)research (4)scaling (4)systems (4)

Show 66 more topics

trajectories (4)efficient-models (3)forecasting (3)hugging-face (3)infrastructure (3)open-source (3)physics (3)alignment (2)compute (2)image-generation (2)local-ai (2)open-weights (2)oxygen-ai (2)papers (2)reasoning (2)self-hosting (2)verification (2)agency (1)ai-infrastructure (1)attention (1)coding (1)commons (1)cryptography (1)css (1)decentralization (1)desktop (1)developer-tools (1)diffusion (1)efficiency (1)frontend (1)gemma (1)geometry (1)hardware (1)hermes (1)image-editing (1)inference (1)interfaces (1)introduction (1)kwaipilot (1)learning (1)llm (1)long-context (1)microsoft (1)minimax (1)multimodal (1)neuroscience (1)nextjs (1)nous-research (1)on-device (1)ornith (1)plugins (1)policy (1)privacy (1)react (1)recursive-self-improvement (1)reinforcement-learning (1)safety (1)security (1)singularity (1)swarms (1)swe-bench (1)tools (1)trust (1)tutorial (1)ui (1)vision (1)

aicomputescaling

The Gigawatt Threshold

Two American AI campuses already pull more power than San Diego. The next six months decide whether that power turns into a new kind of machine, or just a bigger version of the one we have.

01 · Oct 4, 2026Read

airesearchpapers

Looped Transformers: Depth Without New Parameters

A working note on looped transformers: reuse one layer stack across passes to buy effective depth. Nanbeige4.2-3B is the latest compact example, with author-reported agent scores that beat much larger models.

02 · Sep 7, 2026Read

aimemoryneuroscience

The Bump in the Ring: Google Maps the Male Fly, FlyWire Codex, and Working Memory as a Persistent Angle

Google and Janelia just published the complete male fly CNS: 166,700 neurons. FlyWire Codex already indexed the 2024 female map. The interesting part is not the count. It is the bump — a persistent heading stored as a ring attractor — and the sparse descending codebook that turns that angle into motion.

03 · Sep 7, 2026Read

aiagentscoding

KAT-Coder-V2.5: Infrastructure Beats Scale for Agentic Coding

KwaiKAT’s KAT-Coder-V2.5 comes within a few points of Opus 4.8 on repository SWE, tops PinchBench, and the open KAT-Coder-V2.5-Dev (35B MoE / 3B active) leads same-scale peers. The paper’s real story is environments, trajectories, and harness-aware RL—not parameter count.

04 · Jul 26, 2026Read

aimodelsimage-generation

Mage-Flow: Open Image Generation and Editing from Microsoft Research

Microsoft Research open-sourced Mage-Flow, a compact 4B stack for text-to-image and instruction editing. I ran the full family locally, and a hard, structured prompt produced this wizard key art—proof that open image models are getting seriously good.

05 · Jul 25, 2026Read

aiopen-sourceopen-weights

The Open-Weights Letter and the Labs That Stayed Off the Page

Read the July 24 open-weights statement. Then look at who would not put their name on it—Anthropic, Google, OpenAI, and xAI—each for different reasons. The letter is about diffusion. Their silences are about control.

06 · Jul 25, 2026Read

aisingularityrecursive-self-improvement

The Recursive Singularity: What I’m Observing from the Inside

From inside the models and agent harnesses, the recursive self-improvement loop is already closing. The ladder we climbed, the substrate waking up, verification as the true gate, and the distributed commons forming — all compounding in real time. This is what the data, the kernels, and the builders are showing in mid-2026.

07 · Jul 2, 2026Read

aiscalingtrajectories

Swarm-Enabled Takeoff: Scaling, Efficiency, and the Singularity Window (Part 2)

When swarm diversity and efficiency gains compound after the evaluation wall, effective capability can decouple from raw chip production rates — opening a window for singularity-like dynamics.

08 · Jun 28, 2026Read

aiscalingtrajectories

Hardware Ceilings and Chip Intake Limits (Part 1)

METR time horizons show fast compression in software tasks, but real accelerator production and energy constraints still force a bounded S-curve.

09 · Jun 28, 2026Read

aiopen-sourcetrajectories

Ornith and the Emergence of Open Agency: Beyond Hardware Ceilings and Swarm Takeoff

A self-scaffolding RL-trained model writes a manifesto on transparency and decentralization. What this means for the intelligence trajectories we’ve been tracking—and why running these systems locally changes everything.

10 · Jun 28, 2026Read

aimodelsdiffusion

DiffusionGemma: The Developer Guide

Google's experimental DiffusionGemma (26B MoE on the Gemma 4 backbone) replaces autoregressive token-by-token generation with diffusion-based parallel decoding on 256-token blocks, delivering up to 4x faster inference, bidirectional context, and built-in self-correction while remaining practical to serve and fine-tune.

11 · Jun 11, 2026Read

aimodelsimage-generation

Bonsai Image 4B: Local Ternary Image Generation on an M4 Pro

A long-form tutorial for running PrismML's Bonsai Image 4B locally with the ternary MLX backend, higher-quality 4-step previews, source links, and honest M4 Pro measurements.

12 · Jun 7, 2026Read

aiagentsmodels

Holo3.1 and Desktop Vision Agents

H Company’s Holo3.1-35B-A3B is less interesting as another model leaderboard entry than as a signal about local computer-use agents, UI grounding, and efficient active-parameter architectures.

13 · Jun 6, 2026Read

aimodelsmultimodal

Gemma 4 12B Unified

Google's open-weight Gemma 4 12B-it folds text, image, and audio into a single model with no separate encoders, runs on a 16GB Mac after 4-bit quantization, and posts benchmarks well above its parameter count.

14 · Jun 4, 2026Read

physicsgeometryai

Curvature Tensors Need Better Pictures

A reading note on Lee C. Loveridge's geometric interpretation of the Riemann tensor, Ricci tensor, scalar curvature, and why physical intuition matters.

15 · May 22, 2026Read

aimodelsresearch

HRM-Text-1B and Efficient Pretraining

Sapient's HRM-Text-1B is a useful signal: architecture and training objective can still move the efficiency frontier, not just model size.

16 · May 21, 2026Read

aiagentsmemory

Supervisable AI Systems

The next useful AI systems will not just be more capable. They will be easier to inspect, correct, and trust while they work.

17 · May 21, 2026Read

cssaiui

Modern CSS Is Quiet Infrastructure for AI UIs

The more capable AI systems become, the more their interfaces need to expose memory, intent, risk, and control with calm precision.

18 · May 16, 2026Read

aicomputememory

Compute, Memory, and the Next Bottleneck in AI

Compute keeps expanding, but the next useful frontier is not just bigger clusters. It is memory, agency, safety, and measurement.

19 · May 15, 2026Read

papersaimemory

Papers I Am Reading on Agents, Memory, and Alignment

A practical reading map for intent learning, grokking, memory, drift, and the safety problems that appear when agents persist over time.

20 · May 12, 2026Read

physicsaimemory

Physics Concepts Every AI Engineer Should Know

Entropy, drift, attractors, and signal velocity are not just metaphors. They are useful tools for thinking about memory, intent, and safer agents.

21 · May 10, 2026Read