AgentsNov 3
Image: newsletter.pragmaticengineer.com
Tano engineer builds affiliate platform and migrates agents with AI
Using Claude Code, Tano engineer Neil Kakkar built an affiliate platform and migrated the company’s agent architecture, according to his account. His workflow runs coding agents in parallel and uses an AI reviewer on pull requests (proposed code changes). He describes production errors that needed subsequent fixes. Kakkar reports merging about 11 times as many pull requests per month as at his previous employer.
AgentsNov 3
Image: cloud.google.com
Google and Mercor release 29 engineering tasks for AI agents
Google and Mercor released 29 public tasks from an 85-task test of agents doing physical-engineering work through project files and engineering constraints. Agents helped assemble 41 tasks that practicing engineers then validated. Google says the best tested configuration averaged 48.3% on its private set. On tasks written with agent assistance, only 41.3% of that configuration’s deliverables passed every mandatory safety and feasibility check.
ImageNov 4
Image: linum.ai
Linum reports training an image model for about $21,000
For its experimental Pyramid-JiT image model, Linum reports $21,038 in computing costs for initial training and follow-up training. The model has roughly 2 billion parameters (adjustable numerical settings) and works directly with image pixels. Linum published 249 samples and an account of the training run. The team says Claude helped build the gallery and diagrams and prepared the model documentation and repository materials.
Compute & InfraNov 3
Image: blogs.nvidia.com
Lam Research and Advantest report faster chipmaking work with AI
A model from Lam Research predicts the behavior of plasma (the electrically charged gas used to etch chips) in milliseconds, according to NVIDIA. A full simulation takes days, so faster predictions could help engineers examine more process choices before costly physical trials. Advantest also uses AI agents to select models on its chip-testing floor, NVIDIA says. NVIDIA puts Advantest’s response times at minutes, down from hours.
Compute & InfraNov 4
Image: qualcomm.com
Qualcomm reports lab validation of memory-stacked AI chip
Qualcomm says its first High Bandwidth Compute chip passed lab validation, using a design that moves AI computation closer to memory. The design bonds stacked LPDDR (low-power memory) directly to an AI-processing base chip. Early tests showed 18 times the effective bandwidth (data moved per second) of LPDDR5X memory. Qualcomm also reports more than five times the bandwidth per watt of HBM4, another memory technology.
Compute & InfraNov 3
Image: theregister.com
Revolut reports lower AI coding costs on its own hardware
About 2,000 Revolut engineers use GLM models running on the company’s own hardware by default. Revolut says the cost of running its AI coding models fell roughly 90% after the switch. It serves GLM-5.3 on 56 NVIDIA B200 graphics processors, retaining other models for tasks that require image understanding. After its AI-tool rollout, Revolut reports improvements of 60% in coding time and 70% in lead time (the time it takes work to reach delivery).
RobotsNov 2
Image: The Threshold Report/GPT Image 2.5
Researchers demonstrate Pixel wheeled humanoid for movement and manipulation
Researchers demonstrated Pixel, a low-cost wheeled humanoid that combines movement and object handling in one research system. Its design brings together compliant grippers (grippers that yield under contact), human and robot demonstrations, whole-body control and learning guided by language instructions. Researchers tested navigation, dexterity and coordinated movement, along with active perception (taking actions to gather information).
RobotsNov 4
Image: The Threshold Report/GPT Image 2.5
LiberAI reports Liber-0 results on 60 GPT-6-proposed robot tasks
Without robot demonstrations collected specifically for the tasks, LiberAI’s Liber-0 succeeded in 36% of 600 trials across 60 tasks proposed by GPT-6, the company says. The team also tested whether human recordings could reduce the robot-specific data needed to teach object manipulation. Combining human glove recordings with one-quarter of the robot demonstration pool earned a 75.6 progress score, versus 72.8 using the full robot demonstration pool alone.
TextNov 3
Image: cloud.google.com
Nammu21 tests AI-extracted loan terms on Google Cloud’s ledger
Nammu21 and Google Cloud modeled a $1.945 billion term-loan segment to test a shared, programmable version of a credit agreement. Nammu21’s AI system structured the agreement’s terms and turned them into programmable contracts on Google Cloud Universal Ledger. The demonstration modeled participant consent, transfers and repayments against those terms, testing whether parties could avoid repeatedly reconciling separate interpretations.
TextNov 3
Image: The Threshold Report/GPT Image 2.5
IBM trains Palette to make presentations in its design style
IBM Research trained Palette to make presentation decks in IBM’s design style, testing a model that an organization could run itself. The model has 20 billion parameters (learned numerical values) and was trained using about 1,000 AI-generated decks. Its workflow separates planning, slide generation and a checking stage that repairs layout problems. IBM reports scores of 59.7 for Palette and 60.8 for NotebookLM across three presentation categories in the PresentBench test; Palette ranked first in two.
