Drew's Comic NewsroomAbout

Comics on the day’s news

Claude spent the weekend fixing other models’ bad habitsCoyote vs. Acme opens at 95% on the TomatometerAn ETF is a basket that trades like a stockSome quantum-gravity clues may be two maps of the same hillSara Bareilles turns a Lincoln Center lobby into a grief archiveBáez saved it, McGonigle ended it, Tigers 2-1A US Open guard’s wife weaponizes Tsitsipas quotesThe book nobody can read is still unreadClaude spent the weekend fixing other models’ bad habitsCoyote vs. Acme opens at 95% on the TomatometerAn ETF is a basket that trades like a stockSome quantum-gravity clues may be two maps of the same hillSara Bareilles turns a Lincoln Center lobby into a grief archiveBáez saved it, McGonigle ended it, Tigers 2-1A US Open guard’s wife weaponizes Tsitsipas quotesThe book nobody can read is still unread

Technology and AI

Technology and AI

Comics

Four-panel full-color AI comic. Panel 1: cartoon robot stuck choosing the next word greedily. Panel 2: glowing causal topology matrix plugged into a diffusion cloud. Panel 3: intermediate thought bubbles getting revised by arrows. Panel 4: cartoon Sudoku grid lighting up with green checks.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

CaLR: Causal Latent Revision aims to fix diffusion reasoning’s soft logic

arXiv 2609.20981 (Sept. 17): Causal Latent Revision mixes causal topology from an expert model with diffusion language models for gradient-guided thought revision; SOTA DLM results, strong on Sudoku-style constraints.

September 21, 2026 at 7:00 AM PDT

Four-panel full-color AI research comic. Panel 1: cartoon robot under AI MATCH banner. Panel 2: five-step curriculum ladder from foundations to mastery. Panel 3: three glowing gated-reward doorways. Panel 4: scoreboard showing 72.19 percent and 62.87 percent — verified as API-Bank and BFCL V3, not student grades.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

MATCH tool-learning paper posts 72.19% API-Bank, 62.87% BFCL V3

arXiv 2609.20082 submitted Sept. 17: Shihao Liu et al.; Model-Aware Tool Learning with Curriculum Scheduling and Hierarchically Gated Rewards; MACL + HTGR closed loop. Abstract only.

September 20, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: bright cartoon five-finger robot hand. Panel 2: hand walking on fingertips across tile and grass. Panel 3: hand flipping upright after a tip-over. Panel 4: hand pressing keyboard keys while balancing.

Comic: Topics / Drew’s Comic Newsroom. Source: Tech Times / ETH Zurich Soft Robotics Lab.

Technology and AI

ETH Zurich teaches a robotic hand to walk on its own fingers

Tech Times / Soft Robotics Lab, Sept. 18: arXiv ‘Fingers as Legs’; 818 g WUJI Hand crawls 14 surfaces, recovers from falls, presses keys while self-supported.

September 19, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: grid of cartoon robot researchers labeled 37,000 agents. Panel 2: catalog of 50,000 trials. Panel 3: green ON/OFF gene switch. Panel 4: pink-and-blue antibody-drug conjugate cartoon.

Comic: Topics / Drew’s Comic Newsroom. Source: Stanford Medicine.

Technology and AI

Stanford’s virtual biotech fields 37,000 AI agents on drug discovery

Stanford Medicine / Science, Sept. 17: Zou & Zhang’s agent company scored 50k trials; switch-like targets +40% Phase 1→2; B7-H3 ADC design later matched in the wild. Cartoon only.

September 18, 2026 at 7:00 AM PDT

Four-panel full-color AI research comic. Panel 1: cartoon masked sequence being iteratively denoised. Panel 2: position-oriented directed acyclic graph binding node groups to fixed output positions. Panel 3: comparison board showing DA-DLM ahead of Block Diffusion especially with fewer denoising steps. Panel 4: parallel generation stamp noting match to autoregressive quality.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

DA-DLM adapts DAG dependency modeling to diffusion LMs

arXiv 2609.15070: position-oriented DAG for diffusion language models; outperforms Block Diffusion especially with fewer steps; matches AR while staying parallel. Abstract only.

September 16, 2026 at 7:00 AM PDT

Four-panel full-color AI research comic. Panel 1: cartoon attention sinking into the first token position. Panel 2: causal-mask grid illustrating self-concentration and Value-non-mixing. Panel 3: activation spike at the initial position challenging low-bit quantization. Panel 4: EMNLP 2026 paper board on attention-sink mechanisms.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

EMNLP 2026: Attention sinks tied to self-concentration, not RoPE

arXiv 2609.09085: Attention Sink and Massive Activations at the initial position; causal mask + Value-non-mixing drive self-concentration; implications for quantization. Abstract numbers only.

September 15, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: robot cycling a looped hidden-state recurrence. Panel 2: local denoising objectives with decreasing noise levels. Panel 3: 58.8 percent ARC-AGI-1 test accuracy scoreboard. Panel 4: 12.2 percent ARC-AGI-2 badge across six reasoning benchmarks.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

Looped Flows hit 58.8% on ARC-AGI-1 with local denoising recurrence

arXiv 2609.11801, submitted Sept. 10: Thinking with Looped Flows trains looped models with local denoising objectives; 12.2% on ARC-AGI-2 across six reasoning benchmarks.

September 11, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: overwhelmed robot in KV cache library. Panel 2: Declarative Attention tags. Panel 3: skipping shelves, 52% fewer attended tokens. Panel 4: modest 1.27pp accuracy dip scoreboard.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

Declarative Attention lets models skip most of the KV cache

arXiv 2609.02737: zero-shot DA cuts attended tokens ~52% on Gemma-4-31B across 15 long-context tasks with a ~1.3-point accuracy dip.

September 9, 2026 at 7:00 AM PDT

Four-panel full-color AI research comic. Panel 1: a small cartoon language model rubber-stamping YES on a stack of 1200 logic cards. Panel 2: a colorful probe meter reading the correct verdict at 0.96 AUC from hidden states. Panel 3: a giant decision dial offset by plus 4.6 sigma with margin AUC 0.89. Panel 4: three labeled bins reading concealed, miscalibrated, and undetected beside repair dials lifting accuracy from 50 to 81 percent at 0.6B and 94 percent at 8B.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

Internal probes beat the answer: a 0.6B model says YES to everything

Villuri, Shaik, and Doboli show hidden-state probes hit 0.96 AUC while behavior collapses to a +4.6σ threshold offset—method story on miscalibration, not a jailbreak guide.

September 8, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: cartoon transformer looping middle layers and forgetting. Panel 2: Loop Memory Attention sticky notes. Panel 3: halting head traffic light. Panel 4: MathQA 56.9% at about two loops.

Comic: Topics / Drew’s Comic Newsroom. Source: arXiv 2609.03379.

Technology and AI

RecurTrace: looped AI that remembers and knows when to stop

arXiv, Sept. 3: Loop Memory Attention plus an oracle-distilled halting head lifts MathQA to 56.9% at about two loops, beating fixed-depth baselines.

September 6, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: cartoon parchment with a^n+b^n=c^n crossed out for n>2. Panel 2: colorful agent bots building Lean proof bricks. Panel 3: green computer-checked stamp. Panel 4: 11-day scoreboard with theorem count.

Comic: Topics / Drew’s Comic Newsroom. Source: Anthropic, Sept. 4, 2026.

Technology and AI

Claude formalizes Fermat’s Last Theorem in Lean—end to end

Anthropic says an 11-day, largely autonomous run produced the first complete computer-checked FLT proof: millions of lines of Lean and tens of thousands of lemmas.

September 5, 2026 at 7:00 AM PDT

Four-panel full-color AI comic. Panel 1: cartoon robotaxi and confused safety rider. Panel 2: black-box brain wrapped with concepts approaching stopped vehicle and close to cyclist. Panel 3: emergency brake kick-in explained. Panel 4: Nature paper clipboard, humans predict AV mistakes better.

Comic: Topics / Drew’s Comic Newsroom. Source: MIT News, Sept. 2, 2026.

Technology and AI

MIT’s CW-Net helps humans predict self-driving mistakes

MIT News, Sept. 2: a Concept-Wrapper Network translates an AV planner’s opaque decisions into concepts like approaching stopped vehicle — and Nature published the result.

September 4, 2026 at 7:00 AM PDT

Four-panel full-color lab-gallery comic, no real artist portraits. Panel 1: a cartoon detective asks whose art a pink diffusion printer made. Panel 2: removing one painting from a huge pile, the print stays the same. Panel 3: a folder labeled artist removed, still the same print, banner attribution decay. Panel 4: a graph, dataset size up, counterfactual radius down.

Comic: Topics / Drew’s Comic Newsroom. Source: MIT News, Aug. 18, 2026.

Technology and AI

MIT: at scale, AI images often have no one author

CSAIL’s attribution-decay study finds that as a diffusion model’s training set grows, deleting any one image — or even one artist’s whole folder — often leaves the output unchanged.

September 2, 2026 at 7:00 AM PDT

Four-panel full-color lab comic. Panel 1: a cartoon scientist holds a green Nature’s Recipe Book and a yellow notebook labeled new sequences. Panel 2: orange and blue amino-acid characters shake hands with a gold pairwise-energy spark. Panel 3: glittery noise around a protein fold. Panel 4: a smiling cartoon protein ribbon stamped stable fold, not a copy.

Comic: Topics / Drew’s Comic Newsroom. Source: MIT News, Aug. 27, 2026.

Technology and AI

MIT’s PottsMPNN designs proteins without copying nature’s recipes

A Biology Department framework folds physical energy into sequence design so new proteins can be stable without reproducing the amino-acid strings evolution already picked.

September 1, 2026 at 7:00 AM PDT

Four-panel AI comic: Claude at a literature desk, a student model on a training bench, a scoreboard where the machine beats researchers who had eight hours, and a smaller Claude patching a larger successor with a short stack of examples.

Comic: Topics / Drew’s Comic Newsroom. Source: Anthropic, Aug. 28, 2026.

Technology and AI

Claude spent the weekend fixing other models’ bad habits

Anthropic: automated alignment researchers closed safety gaps on ten failure types, and the methods still worked on models up to 4.7 times larger.

August 30, 2026 at 7:00 AM PDT

Four-panel comic: a student with an answer key and a 100-star score marked mush, two locked boxes labeled benchmark and model, two people who cannot see each other’s secrets, a small Gemini Flash Lite robot taking a hidden quiz.

Comic: Topics / Drew’s Comic Newsroom. Source: Google DeepMind, Aug. 27, 2026.

Technology and AI

If the model already saw the test, the score is mush

DeepMind is piloting a double-blind evaluation of Gemini Flash Lite so the model and the benchmark stay hidden from each other.

August 29, 2026 at 7:10 AM PDT

Four-panel comic. A tech loads a zipper demo; the robot half-opens a pouch at ~59%; it improvises with a banana; two clips in context.

Original comic, AI desk. News: Generalist AI / TechTimes.

Technology and AI

Show the Robot Once. It Tries the Zipper.

Generalist AI published GEN-1.5, a robot foundation model that can pick up a short manipulation task from a single 3-to-12-second demonstration, with no gradient updates. “Physical prompting” into a 30-second context window; 100 Hz actions. About 59% average one-shot success across 10 short tasks, about 83% after 10 gradient steps on five minutes of data. Emerged after more than eight months of pretraining. Can chain two demos; sometimes improvises tools. Commercial focus remains GEN-1.

August 22, 2026 at 3:10 PM PDT

Four-panel comic. A shopper finds a computer in a Waymo trunk; close-up of the core; Engine A and Engine B; groceries stay in the back seat.

Original comic, AI desk. News: Waymo / The Verge / TechRepublic.

Technology and AI

Waymo Puts a Data-Center Brain in the Trunk

Waymo opened the trunk on the computer that drives its robotaxis. A purpose-built 5-nanometer ASIC cleans camera, lidar, and radar streams, runs temporal denoising, and fuses sensor data. Those ASICs alone deliver more than 1,000 TOPS. Onboard capacity has grown twentyfold in eight years. Two independent engines; no human backup. TechRepublic: silicon entering production in the Ojai (Geely-owned Zeekr); service on the order of 500,000 paid trips a week. Paired with CPUs, GPUs, and partners including AMD, Nvidia, and TSMC.

August 22, 2026 at 3:00 PM PDT