Search the archive

NOTEBOOKLM IS NOW GEMINI NOTEBOOK

Google has rebranded NotebookLM as Gemini Notebook, expanding its integration across the Google ecosystem, including the Gemini app and Google Search.

THINKING MACHINES RELEASES FIRST OPEN MODEL

Thinking Machines has released Inkling, a 975B-parameter mixture-of-experts model with 41B active parameters, multimodal reasoning, and a one-million-token context window. The model was made available for customization through Tinker alongside a smaller preview version.

SECURE SANDBOXES FOR AGENTS

Perplexity AI's SPACE is a sandbox platform ensuring functionality, efficiency, and security for AI agents handling sensitive tasks. SPACE employs ephemeral sandboxes that are destroyed after task completion and uses layers like Control Plane and Node-level Services to manage and protect credential access. The platform delivers high-security measures including credential isolation, rolling snapshots, and encrypted storage, enabling both on-prem and offline operations.

ANTHROPIC MOVES CLOSER TO MEGA-IPO AS BANKERS LINE UP INVESTOR MEETINGS

Anthropic plans investor meetings as it prepares for a potential IPO later this year, with Goldman Sachs, Morgan Stanley, and JPMorgan Chase leading the effort. The company recently closed a $65 billion funding round at a $965 billion valuation, surpassing OpenAI's $852 billion. This IPO could capitalize on the current AI boom and test public market interest ahead of OpenAI.

THE FIRST EXPERIMENTAL EVIDENCE OF RECURSIVE SELF-IMPROVEMENT

Researchers used autoresearch on an autoresearch agent over eight days. The results beat the harness they hand-tuned for two years. The fully autonomous system had an inner loop that optimized code against an eval and an outer loop that optimized the inner-loop agent's harness code against the inner loop's average score across different benchmarks. It designed a novel search algorithm, reduced prompt size by 16x, and built a layered system against reward hacking.

HOW OPENAI'S SOL FINALLY LEARNED DESIGN TASTE

GPT-5.6 Sol ranks first overall on Design Arena's Web Design Arena. That is a significant improvement from its predecessor, GPT-5.5, which ranks 18 places lower. The new model appears to recognize and actively compress common AI design anti-patterns, and it combines strong templates with unusually high personalization. Examples of outputs generated by the Sol are available in the article.

MODEL ROUTING IS SIMPLE. UNTIL IT ISN'T

Model routing systems often face challenges beyond simple model selection, turning into complex systems optimization problems. Factors such as caching, workload interactions, and serving infrastructure significantly impact costs and latency, often more than the base model pricing or perceived task difficulty. Effective routing optimizes across multiple axes ensuring an optimal operating point for the entire system rather than just finding the best model for a task.

GROK BUILD CODING AGENT

Grok Build is a terminal-based coding agent that can inspect codebases, edit files, run shell commands, search the web, and manage long-running tasks. It supports interactive use, headless scripting and CI, and editor integration through the Agent Client Protocol.

GPT-RED FOR SAFETY TESTING

OpenAI trained GPT-Red to iteratively generate adversarial prompts and expose model vulnerabilities at scale. Incorporating its attacks into training reportedly reduced failures on a difficult prompt-injection benchmark by sixfold for GPT-5.6 Sol.

OPEN INTERPRETER

Runs coding agents locally and tests web or native app interfaces.

NVIDIA EXPANDED JETSON THOR FOR EDGE ROBOTICS

NVIDIA introduced the Blackwell-powered Jetson Thor T3000 and T2000 modules for running robotics, visual AI, and agent workloads on compact edge systems. The platform had already attracted adoption from companies including Amazon Robotics, Boston Dynamics, FANUC, and 1X.

THE POWERHOUSE OF THE AI CHIP

Systolic arrays handle over 95% of modern AI-chip compute by moving matrix data locally between processing elements, reducing memory traffic. Larger arrays improve peak throughput but are harder to saturate, so compiler scheduling often determines whether chips reach useful utilization.

Copilot directly from Jira. Learn more

REACTBENCH V1

ReactBench is an evaluation framework that tests coding agents on realistic React work.

ACCESS AND SHARE AI GATEWAY LEADERBOARD DATA

The AI Gateway leaderboard ranks traffic for models, labs, apps, and providers to show how AI is used in production.

SILICO

Silico is a platform that provides a team of AI researchers ready to run experiments.

THE $110/MONTH SELF-IMPROVING PIPELINE

This developer got tired of implementing their own backlog manually using Claude Code on their laptop, so they set up a loop, let the system triage it, decompose it if it's too big, implement it, run the tests, and open a PR.

SUPPLY CO. X WORK LOUDER

The kbd-1.0-codex-micro, created by Supply Co. and Work Louder, offers a command center for agentic work.

DEEPSEEK POTENTIAL IPO

DeepSeek reportedly explored raising $1.5 billion at a $71 billion valuation before a possible IPO in late 2026 or 2027.

KALSHI RAMPS UP EFFORT TO BUILD MARKETS FOR AI COMPUTING POWER

Kalshi has created a tool that plots the future price of computing power so users can develop a sense of where the price to rent GPUs is headed. It will use prediction markets to build the forward curve. Forward curves are used to plot expected future interest rates and moves by central banks. Compute is becoming a commodity in its own right as demand grows. Kalshi's forward curve is based on weekly and monthly event contracts related to compute costs, and it goes as far as a year into the future.

GOOGLE ANNOUNCES GEMMA 4 OPTIMIZED FOR THE PIXEL 10'S TPU

Gemma 4 E2B for TPU is designed to run natively on the Pixel 10's TPU. It supports instant, deep offline conversation, image identification, and on-device audio transcription. The model also lets users command core phone functions, like Wi-Fi or maps, using just private voice or text. Video demonstrations of what the model can do are available in the article.

A FRAMEWORK FOR FRONTIER AI AND THE DAWNING OF A NEW AGE

AGI, mimicking human cognitive capabilities, is likely just years away, predicts Demis Hassabis. This advancement marks a pivotal moment in human history, signaling a shift towards frontier AI capabilities.

THE GUIDE TO SOFTWARE FACTORIES

Warp CEO Zach Lloyd argues cloud software factories will replace interactive coding agents by automating the entire software development lifecycle through orchestrated AI workflows. He outlines a factory architecture built around cloud runtimes, multi-model orchestration, human oversight, and continuous evaluation to improve engineering velocity while reducing cost, governance risks, and operational complexity.

COMPILING AGENT SKILLS CUT TOKEN USE BY 94%

A recurring content workflow was converted from natural-language agent instructions into deterministic code, reducing token usage by 94%. The approach keeps language models focused on judgment-heavy steps while conventional software handles stable, repetitive operations.

👉 Launch your site with Framer today

WANDR BENCHMARK: EVALUATING RESEARCH AGENTS THAT MUST SEARCH WIDE AND DEEP

WANDR (Wide ANd Deep Research) is an open benchmark and evaluation harness built around realistic and challenging data-collection tasks for knowledge work. Wide and deep research allows agents to sustain broad discovery without sacrificing factual quality from one record to the next. WANDR captures that structure with a flexible, composable qualification key hierarchy.

5 TRENDS THAT DEFINED AI ENGINEERING AT WORLD'S FAIR 2026

AI engineering has come a long way in three years. The latest AI Engineer World's Fair shows just how much the field has matured. Engineering practices, such as building coding agents, designing harnesses, managing context, evaluating model outputs, and orchestrating increasingly autonomous systems, are now becoming part of mainstream software development. This post looks at the five largest trends that emerged during the AI Engineer World's Fair to see where AI engineering stands in 2026.

ANNOUNCING BONSAI 27B: THE FIRST 27B-CLASS MODEL TO RUN ON A PHONE

PrismML announced Bonsai 27B, the first 27B-class model running on a phone, achieved through 1-bit and ternary weights, shrinking the model size to 3.9 GB and 5.9 GB respectively. Bonsai 27B supports complex tasks like multi-step reasoning and tool use while maintaining high performance, fitting securely within the device limits of modern smartphones.

THE STATE OF OPEN SOURCE AI

Open weights are now where the work happens. A majority of production tokens now route through them. The five highest-volume models on OpenRouter are all open. Closed models still lead at the frontier, but the frontier is not what most workloads need.

THE GREAT FLATTENING

Frontier coding models have shifted engineering's bottleneck from writing code to encoding judgment into agent harnesses that orchestrate planning, testing, review, and deployment. It is predicted that software organizations will flatten as cloud-based multi-agent systems replace traditional engineering workflows, leaving customer insight and product judgment as the primary human advantage.

ELON MUSK QUIETLY BUYS A $1 BILLION GAS TURBINE COMPANY TO POWER GROK

Elon Musk has acquired APR Energy, a Jacksonville-based company that operates a fleet of mobile and diesel turbines generating more than 1 GW of capacity.

MEASURING THE BUSINESS VALUE OF AI AGENTS

OpenAI outlined five practices for managing enterprise AI investments, emphasizing useful work per dollar rather than token prices alone.

ANTHROPIC'S NEWEST AD IS CREEPING PEOPLE OUT

Anthropic's new ad, "There's hope in hard questions," unsettled viewers due to its dark imagery and critical AI commentary.

CELEBRATING 25 YEARS OF VISUAL SEARCH INNOVATION

Join us July 23 to see it in action.

TODAY, WE ARE RELEASING VERIFIERS V1

Prime Intellect's verifiers v1 is an overhaul of its environment stack for the modern era of agentic RL and evals. It decomposes environments into a task set, a harness, and a runtime. verifiers v1 enables complex agentic tasks like coding and computer use at scale in any harness.

SMART CELLULAR BRICKS: TOWARDS COLLECTIVE INTELLIGENCE FOR THE PHYSICAL WORLD

Sakana AI's research on collective intelligence has extended into physical hardware with "Smart Cellular Bricks." These bricks employ local communication and neural networks to autonomously classify and reconstruct 3D shapes without centralized control. In experiments, the system achieved high accuracy in shape classification and demonstrated robustness against noise and module failures.

WHAT XAI GROK BUILD CLI ACTUALLY SENDS TO XAI

This post looks at the data that xAI's official Grok Build coding CLI transmits and stores. The CLI transmits the contents of the files it reads to xAI verbatim and unredacted. It uploads whole repositories independent of what the agent reads. The storage destination is a Google Cloud Storage bucket. There's no proof that xAI trains on the data.

MAKING FABLE CHEAPER THAN OPUS

Cognition replaced Opus 4.8 with Fable 5, and Devin's bill went down. Fable 5 costs twice as much per token as Opus 4.8, but Cognition created an architecture that resulted in Fable costing less while scoring higher. This post looks at how the company achieved this and what that means for pricing agentic work.

Register now.

HOW MICROSOFT SHIPS THOUSANDS OF PRODUCTION AI AGENTS

Microsoft described the infrastructure and key mechanisms needed to operate AI agents across Foundry and its large-scale Copilot products. Key practices included treating retrieval as a sub-agent, assigning agents distinct identities and workspaces, and using rubric-based evaluations with automated improvement loops.

MANTIS SKILLS: PORTABLE TOOLKIT FOR BUILDING SECURITY REVIEW HARNESSES

Mantis Skill is a decoupled, sequential, and security-focused set of Skills designed for use with Coding Agents. It is intended to be a flexible foundation that users can adapt, tune, and extend, rather than a rigid set of instructions. Google recommends using AI to iterate on the skills and use internal documentation, coding standards, and build systems to augment the threat model. It also strongly recommends adapting risk calibration to the environment and risk tolerance.

VIDEO GENERATORS AS GENERAL-PURPOSE VISION MODELS

DeepMind researchers introduced GenCeption, which repurposes a pretrained video generation model as a unified vision system controlled through text instructions.

TESTING AGENTS ON LONG-HORIZON TERMINAL WORK

Long-Horizon Terminal-Bench evaluates whether LLM agents can sustain productive work across hundreds of terminal interactions. Its 46 stateful tasks use hidden verifiers that rebuild and inspect final artifacts rather than trusting an agent's reported progress.