AI Breakthrough Tracker

Cal Poly's new AI Factory gives students and faculty direct access to advanced AI computing on campus.

• Hardware and funding: The facility uses...

A small team rarely has spare reviewers, so the agent's output can become the bottleneck. Three limits matter most:

• Tight tickets: A vague request...

Robots are learning fine manipulation from two directions. Simulation serves as the practice ground, with researchers now training robots on tasks...

DentScribe reports that its engine passed all 68 tests, but the test measures how the engine converts supplied text into chart entries. Read the...

Google says Gemini for healthcare is designed to "unify the disparate clinical, operational and administrative systems", turning fragmented data into...

The Sage-Taiyi architecture monitors and revises its understanding of evidence, assumptions, and unresolved questions, then uses those assessments to...

Cheaper rollouts come with a trade-off. On-policy distillation tolerates substantial training-inference mismatch without meaningful loss of training...

The core tension is that voxel generators have typically meant choosing between high fidelity with heavy compute and light weight with lower...

A simulated 180-day vending benchmark, covering 20 machines, tests agents beyond single transactions:

• Better metrics didn't guarantee growth: Gains...

Internal states and verbal reports can come apart. Across 52,920 reports, pain reports never exceeded what altered text revealed, even though probes...

This paper argues that prompt engineering is an interaction technique, while programmatic prompting is an engineering discipline covering prompts as...

Public anxiety is growing over reports of out-of-control AI agents, while the small federal division that assesses new AI offerings lacks a permanent...

Reproducing an ML experiment means rebuilding steps from details scattered across papers, code, and supporting files. For coding agents, success may...

Retrieval-augmented generation still falls short due to inadequate retrieval and potential hallucinations. The TRIO paper addresses this with three...

A post by @_akhaliq points to LingBot-Touch, described as a foundation model for heterogeneous tactile representation learning. The source is only a title and link, so its methods, data, and robot manipulation results aren't covered here.

The SIRIS paper asks whether language models can separate consequential simulation errors from harmless ones in AI-driven systems research. The source shares only the title and a paper link, so its findings aren't visible here.

Google's November 5, 2026 walkthrough shows a workflow that starts on paper: photograph a hand-drawn sketch, attach it in the Gemini app, turn on...

Both pieces show simulation becoming the place where robot behavior gets tested before it reaches the real world.

• Agents write and test code: AI...

A job-market paper by Chen and Stratton suggests coding agents boost output at the keyboard while review soaks up the difference.

• Code rises,...

Four potential 2028 Democratic presidential contenders met with AI safety advocates and party veterans in San Francisco just days after the midterms....