Search the archive

Explore advice from experts:

and fast AI implementation. Register now

MORE DEVINS IN MORE PLACES

Cognition raised over $1B at a $26B valuation, with significant backing from major investors to expand Devin, an AI software engineer. Devin has significantly cut project times and improved automation for clients like Mercedes-Benz and Itaú. Cognition aims to further streamline software development by matching models to tasks while expanding its engineering capabilities.

ELEVENLABS MUSIC GENERATION MODEL

ElevenLabs has released Music v2, a music-generation model capable of switching genres mid-track while maintaining vocal and compositional coherence.

BIOHUB RELEASES A WORLD MODEL OF PROTEIN BIOLOGY

Biohub has made its open discovery engine for protein structure prediction, design, and biological discovery available to researchers everywhere. The release includes ESMC, a state-of-the-art language model that has internalized the fundamental properties that govern protein biology; ESMFold2, a design engine designed to transform ESMC's sequence representations into atomically-resolved 3D structures of biomolecular complexes; and ESM Atlas, which makes ESMC's representations navigable across 6.8 billion protein sequences and 1.1 billion predicted structures. All three models are freely available to the global scientific community.

SHIPPING A TRILLION PARAMETERS WITH A HUB BUCKET: DELTA WEIGHT SYNC IN TRL

The blog post introduces a method to reduce the weight synchronization payload in async RL using "Delta Weight Sync," which transmits only changed model parameters between RL steps, significantly reducing data transfer from gigabytes to megabytes. A Hugging Face Hub "bucket" manages high-frequency object storage, enabling separate locations for the trainer and inference engine without direct communication, leading to substantial bandwidth savings.

BUILDING SELF-IMPROVING TAX AGENTS WITH CODEX

Real-world systems often behave differently in production than they do in the lab. Teams often discover these failures after launch, then spend weeks fixing them. That feedback loop is slow and manual. Today, it is possible to build agents that self-improve. This post looks at how OpenAI used Codex to build this type of agent at Thrive Holdings, resulting in an AI that can prepare increasingly complex tax returns.

I THINK ANTHROPIC AND OPENAI HAVE FOUND PRODUCT-MARKET FIT

Both Anthropic and OpenAI have started aggressively pricing their APIs. This is likely because they have found product-market fit with coding/general-purpose agent products. Companies spending over $200 per month per user helps these businesses cover their costs much better than charging $10 to $20 per month per user. Coding agents amplify this spending significantly.

Amazon Web Services (AWS). Register now

SECURE MCP TUNNEL

Secure MCP Tunnel enables connecting private MCP servers to OpenAI products without exposing them to the internet. It uses tunnel-client to establish outbound HTTPS paths for request handling while maintaining server privacy. The solution integrates easily with existing systems, supporting enterprise networking requirements and maintaining secure data flow.

INTRODUCING APEX: A FAST, SPECIALIZED MODEL FOR REACT NATIVE

Apex is a React Native coding model trained to build apps by analyzing architecture decisions, fixing framework-specific issues, and reasoning about constraints. While it doesn't match frontier models on coding benchmarks, the optimized model significantly alters the performance-to-cost ratio within its specific domain. The model is still in development. It is now available in a private beta with selected teams.

NVIDIA'S LOCATEANYTHING FOR FASTER GROUNDING

NVIDIA's LocateAnything is a vision-language grounding framework that decodes bounding boxes in parallel rather than token-by-token.

LITEPARSE V2.0

LiteParse is a standalone OSS PDF parsing tool that provides high-quality spatial text parsing with bounding boxes without proprietary LLM features or cloud dependencies. It features fast text parsing, screenshot generation, and support for multiple languages, platforms, and output formats. Everything runs locally on users' machines.

NVIDIA BETS $150B ON TAIWAN AS TRUMP'S PLAN TO MAKE US AN AI HUB BACKFIRES

Nvidia will invest $150 billion a year to make sure that Taiwan remains at the epicenter of the AI revolution. The investment is aimed at cementing Taiwan as the world's tech manufacturing hub for a long time. Nvidia will create a new headquarters in Taiwan to expand its partnership with TSMC, benefit from close proximity to advanced packaging technology not yet available at TSMC's factories in the US, and boost its alliances with other nearby partners. Expanding the AI ecosystem helps Nvidia further its bottom line.

YOUTUBE EXPANDS AUTOMATIC AI VIDEO LABELING

YouTube said it will automatically apply labels to videos containing significant photorealistic AI content, reducing reliance on creator self-disclosure.

data foundations for AI. Register now

EPICURE: NAVIGATING THE EMERGENT GEOMETRY OF FOOD INGREDIENT EMBEDDINGS

Epicure is a family of three sibling skip-gram ingredient embeddings retrained from scratch on a multilingual recipe corpus.

FORMER GOOGLE AND APPLE RESEARCHERS LAUNCH TRAJECTORY TO ENHANCE AI FEEDBACK LOOPS

The new Palo Alto-based startup aims to make AI systems that can actually see and interpret the physical world.

FINDING HIGH-SEVERITY SECURITY ISSUES WITH PUBLICLY AVAILABLE MODELS

Ramp pointed roughly 10,000 Inspect coding-agent sessions at its backend in an 8-hour run with a minimal "find security issues" prompt.

FULLY REMOTE)

TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [26].

GOOGLE DEEPMIND'S HASSABIS: AGI IS 3 TO 4 YEARS AWAY

Google DeepMind CEO Demis Hassabis now predicts AGI could be achieved by 2029-30, accelerating from his earlier estimate of 2030-2035.

GOOGLE EXPANDS GEMINI FOR BUSINESS WITH SHAREABLE PROJECTS

Google enhances Gemini for Business with the introduction of shareable Projects, which allow team members to collaborate within dedicated multi-surface workspaces.

ANTHROPIC TO EXPAND CLAUDE VOICE MODE TO MORE LANGUAGES

Anthropic plans to expand Claude's voice mode to 18 new languages.

subscribers. Grab time with the team to get set up

MUSK'S XAI WARNS STAFFERS TO LIMIT CONTACT WITH CURSOR EMPLOYEES

xAI's top lawyer has warned xAI employees to carefully moderate their interactions with workers from Cursor. Staff should not extend beyond what is necessary to implement a technical partnership. The warning is standard during acquisitions, but it is coming a bit late as the companies' employees have been working alongside each other for weeks. Any accusation that the two sides improperly co-mingled their business could put the acquisition deal in jeopardy.

CHINA EXPANDS TRAVEL CURBS TO TOP AI TALENT AT PRIVATE FIRMS

China has restricted overseas travel for top AI professionals in private firms. These individuals will need approval from relevant authorities before embarking on overseas travel. The restricted individuals include a mix of startup founders, researchers, and executives. China has previously restricted travel for key personnel, from prominent researchers to nuclear scientists and executives at state firms, but it is unusual for the travel restrictions to be extended to private firms.

MAI-IMAGE-2.5 LAUNCHES AT NO. 3 ON ARENA

MAI-Image-2.5 ranks third on Arena's text-to-image leaderboard, excelling in style variety, accurate text rendering, and detailed imagery. It improves significantly over MAI-Image-2 with advancements in visual reasoning, scene structure, and commercial illustration capabilities. These enhancements enhance its ability to transform simple instructions into polished images.

HOW WE CONTAIN CLAUDE ACROSS PRODUCTS

Agents are a new category of software, but their system-level interactions are not. Design for containment at the environment layer first, then steer behavior at the model layer. Match isolation strength to the user's capacity for oversight, and use battle-tested components. AI deployment can be risky, but placing a hard limit on the potential damage often shifts the balance in the right direction.

EXTRACT MORE KERNEL PERFORMANCE WITH NVIDIA COMPILEIQ AUTO-TUNING

NVIDIA's CompileIQ, integrated into CUDA 13.3, uses AI-driven evolutionary algorithms to auto-tune GPU compiler settings, optimizing performance for specific workloads beyond standard heuristics. It provides fine-tuned compiler configurations for individual kernels, offering up to 15% performance improvements on already-optimized AI inference and training tasks. With CompileIQ, developers define optimization objectives, allowing for multi-objective tuning, including trade-offs between runtime, power, and compile time, making it suitable for high-impact applications like LLM inference.

slower than you think. Read the full breakdown from You.com

DEEPSWE

DeepSWE introduces a sophisticated benchmark for long-horizon software engineering, with tasks spanning 91 repositories in 5 languages and ensuring no model has pre-seen the solution. It delivers four key improvements: tasks are contamination-free, reflect real-world complexity, cover diverse repositories, and employ reliable verification processes. DeepSWE provides sharper separation metrics for coding agents, contrasting with the clustering seen in existing benchmarks like SWE-Bench Pro.

NATIVE MULTIMODAL MODELS

This repository catalogs work moving from modular multimodal assembly toward native multimodal modeling, where different modalities are integrated inside a unified transformer space or joint backbone.

OPENROUTER MORE THAN DOUBLES VALUATION TO $1.3B IN A YEAR

OpenRouter raised $113 million in a Series B round led by CapitalG, bringing its valuation to $1.3 billion. The AI gateway startup provides access to over 400 models and processes 100 trillion tokens monthly. OpenRouter's growth reflects a shift in AI towards multi-model solutions, allowing companies to avoid dependence on a single model provider.

SPACEX HAS TWO AI COMPUTE STORIES; ONLY ONE GENERATES REVENUE

SpaceX's S-1 tells two stories. The first is that the company is spending billions building terrestrial data centers and has signed one disclosed external customer, Anthropic, with a deal worth $1.25 billion per month through to May 2029. The second is that the future of AI inference belongs in orbit and that SpaceX is the only company that has already accomplished the key technical challenges associated with evolving connectivity satellites into AI compute satellites. Both stories are presented with conviction, and neither is contingent on the other being wrong.

FULLY REMOTE)

TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [19].

ANTHROPIC TO INTRODUCE AI FLUENCY SCORECARD IN CLAUDE

Anthropic plans to introduce an AI Fluency scorecard in Claude that evaluates user interaction skills across 11 behavioral indicators.

CLAUDE MYTHOS REPORTEDLY SOLVES OPENAI'S LANDMARK ERDŐS PROBLEM WITH A "CUTE, SIMPLE PROOF"

Mythos' solution was a bit worse than OpenAI's, but Mythos was reportedly able to find OpenAI's solution too.

INITIAL RESULTS ON LEGAL AGENT BENCHMARK

Harvey baselined frontier models on its Legal Agent Benchmark holdout under an "all-pass" standard requiring every rubric criterion to pass, and Claude Opus 4.7 led at just 7.1% with Sonnet 4.6 at 5.4%, Opus 4.6 at 4.2%, GPT-5.5 at 2.1%, and Gemini 3.5 Flash at 0.8%, signaling legal work is far from saturated by frontier intelligence.

HOW CLAUDE COWORK'S LEAD ENGINEER USES AI

Felix Rieseberg showed how he used Claude Cowork for tasks like converting a 2D house plan into a 3D floor planner, mining email as a personal inventory database, and building live dashboards from connected apps.

Download the workbook

INTRODUCING GROK BUILD

Grok Build, a new coding agent and CLI, has launched in beta for SuperGrok and X Premium Plus subscribers. It supports complex coding projects by allowing plan mode reviews and integrates seamlessly with user conventions. Users can deploy Grok's capabilities for automation and parallel processing using headless mode and specialized subagents.

NOTES ON POPE LEO XIV'S ENCYCLICAL ON AI

Pope Leo XIV recently released a document on the ethics of integrating AI into modern society. It touches on the environmental impact of the technology, covers the risks of algorithmic systems that make decisions that impact people's lives, discusses how AI amplifies the power of those with resources, and more. A link to the document is available in the article. The writing style is very approachable, even to non-Catholics.

ON AI HARDWARE

The market is becoming a stack of memory problems. Hardware changes slowly, while software and model architectures can move quickly. Hardware companies will need to build architectures that remain useful as the bottleneck shifts.

GEMINI 3.5 FLASH LOOKS GOOD FOR HOW FAST IT IS

Zvi judges Gemini 3.5 Flash the best model at its speed point but unconvincing against Opus 4.7 or GPT-5.5 outside latency-sensitive workloads, with Google positioning it as a daily driver for agentic workflows that outscores 3.1 Pro on Terminal-Bench and MCP Atlas while running 4x faster.