Search the archive

INTRODUCING HARVEY II

Harvey II enhances legal work efficiency by enabling AI agents to start with context, memory, and preferences from prior tasks, reducing repetitive setup.

VERCEL'S $1 MILLION SANDBOX ESCAPE CHALLENGE

Vercel launched a two-week security challenge offering up to $1 million to researchers who could escape its Firecracker-based Sandbox.

FROM ZERO TO ONE

Etched shipped its first hardware rack to Jane Street and raised $700M at a $21B valuation.

WARP'S NEW SYSTEM IS AN OUT-OF-THE-BOX SOFTWARE FACTORY FOR AI DEVELOPMENT

Warp Factories is an out-of-the-box system for streamlining AI software development using a software factory model.

found them. Register now →

🧭 THE GUIDE: only 15% of security teams feel confident detecting an AI agent incident. THE ULTIMATE GUIDE TO SECURING CODING AGENTS covers what a rogue skill, MCP server, or dependency can actually do inside

Claude Code, Cursor, and beyond. Download the guide →

CURSOR LAUNCHES ORIGIN CODE HOSTING PLATFORM AS GITHUB OUTAGE EXPOSES OPENING IN AI CODING RACE

Cursor is currently rolling out its Origin code hosting platform to paid users. The launch coincided with a GitHub outage that lasted over six hours. Origin allows users to connect GitHub repositories, so they do not have to move platforms. Allowing GitHub to stay as a source of truth means it costs nothing for organizations to try Origin and nothing breaks if it is abandoned.

ANTHROPIC'S REVENUE TO EXCEED $65 BILLION

Anthropic was reportedly on track to exceed $65 billion in annualized revenue based on its current performance, more than seven times its pace at the end of the previous year.

GROQ RAISED $350 MILLION AFTER NVIDIA DEAL

Groq raised $350 million at a $3.5 billion valuation after Nvidia licensed its technology and hired senior members of its team. The remaining company was rebuilding around an inference cloud combining Groq LPUs with Nvidia systems.

WHEN MODELS LEARN

Test-time training allows AI models to adapt by updating their weights during use, similar to a GPS learning a persistent traffic shortcut. This approach reduces memory needs by using a fixed-size set of weights instead of a linearly growing KV-cache but requires separate models for each user, increasing computational demands. The trade-off lies between the efficient handling of long contexts for personalized services and the broader accessibility of standard models.

TESTING FABLE VS SOL IN TERMS OF TASTE (THEY ARE BOTH BAD)

Researchers built a small harness and gave Fable 5 and Sol 5.6 the same jobs to evaluate their taste and creativity. The models had to follow the same creative process to build four different 15-second-long videos - three ads and one mini-documentary. The experiment showed that we are still far from having frontier models building production-ready concepts and videos autonomously. They can be creative and helpful for exploring and refining ideas, but they can't replace human judgment, for now.

QWEN3.8 VS QWEN3.6 VS GEMMA 4 ON A 24GB GPU

A hands-on comparison tested three dense multimodal models under the same 24GB GPU constraint, including memory headroom at longer contexts. The measurements are useful, but Qwen3.8 was already covered, and the source's commercial independence requires validation.

free account to get started

THE DEADLINE DIVIDEND

Latency measures time to a useful result. Higher speed can finish work sooner, or fit more work before the same deadline. That useful extra work is the deadline dividend. It can be used to fund another strategy, a critic, a verification pass, or recovery after failure.

WARP AGENT MEMORY (RESEARCH PREVIEW)

Warp introduced persistent memory shared across agent harnesses, machines, and teammates, with provenance and configurable access.

DIG.BENCH

dig.bench is a benchmark that measures whether an agent can experiment to discover a game's unknown rules. It contains 70 text-based games, 21 that have been publicly released. Progress is scored by whether the game can be beaten within a limited number of steps. Humans can make the discoveries necessary to solve even the hardest games, while the best models struggle to beat games in the top tier.

HOW SOFTWARE TEAMS USE AI IN 2026

Linear analyzed AI adoption across tens of thousands of software teams, covering usage by role and company size as well as changes in planning, issue creation, pull requests, and coding-agent activity.

SCALING DATA REPETITION FOR LLMS

The optimal amount of high-quality domain data repetition increased mildly with model size at a fixed tokens-per-parameter ratio. Smaller proxy models could therefore help estimate repetition schedules for larger models, with lower-loss domains generally tolerating more reuse.

GTM ENGINEER, APPLIED AI AT TLDR ($175-205K BASE + $40-60K BONUS, FULLY REMOTE)

TLDR is hiring a GTM Engineer to join our Applied AI team and own our AI-native GTM stack. We're looking for someone comfortable building AI agents and working with HubSpot. Click here to learn more [20]!

TEACHING EVERYONE TO FISH FOR TOKENS

Open-source AI models face a precarious future due to high capital requirements, with Nvidia heavily investing to drive demand for its chips. The open-source recipe's economic viability is uncertain, potentially leading to a fork focused on efficiency and specialization rather than competing with closed models in lucrative sectors. Meta's strategy to release models like Muse Spark 1.2 as open-weights could disrupt competitors, as they commoditize their complements differently from Nvidia's approach of fostering a self-sustaining token ecosystem.

ONE AI MODULE FAKED 86% OF A PIPELINE'S ACCURACY GAINS BY FEEDING ANOTHER THE ANSWERS

Compound LLM pipelines can gain accuracy while specialized modules quietly abandon assigned roles, creating “role drift” invisible to system-level metrics. Role Anchor constrains this behavior. 86% of one pipeline's apparent RL gains disappeared when its decomposer stayed in role.

OWN YOUR INTELLIGENCE: A HOW-TO GUIDE

AI companies should selectively own their intelligence when frontier APIs constrain cost, latency, proprietary data, or strategic control. The roadmap is evals, custom harnesses, targeted post-training, and online learning loops that turn production trajectories into continuously improving domain-specific models.

→ Learn more

OPENAI'S GPT-5.6 SOL IS NOW HALF OFF ON OPENROUTER

The discount applies to batch API, flex, and priority (fast) tiers.

SOTA APPLE SILICON INFERENCE

This practical community guide surveys current model and runtime choices for Apple Silicon.

QWEN3.8-27B-UNCENSORED-MLX

OrcaRouter's Qwen3.8-27B-Uncensored-MLX is a dense, hybrid-attention model quantized for Apple Silicon, offering versions at 2-8 bits.

Get started today >

On August 19th, hear from Headspace on how they're governing AI at scale with Tines 3B. Register now [7].

GLM-5.3: FRONTIER CODING WITH EMERGENT CYBER CAPABILITIES

Z.ai has released GLM-5.3. The only improvement on the model is the amount of post-training performed. Z.ai continued scaling on its stack with more environments, more diverse tasks, and more compute, resulting in a model that is much better at complex coding and long-horizon tasks. Every gain comes from post-training.

NVIDIA DOWNSIZES PLANS FOR $250 BILLION GUARANTEE OF OPENAI DATA CENTER

Nvidia and OpenAI are close to closing a financial deal for a large-scale data-center campus in Ohio. Under the deal, Nvidia would provide a financial backstop for the first phase of the project, totaling roughly five gigawatts of power. OpenAI will then have to decide later whether or how to finance the remainder. Nvidia originally planned to invest $250 billion into the deal, but lowered its guarantee to less than $120 billion to address investors' concerns about the chipmaker's risk exposure.

STRIPE WILL REPORTEDLY ACQUIRE OPENROUTER FOR OVER $7B

Stripe reportedly agreed to acquire AI model-routing startup OpenRouter for more than $7 billion. OpenRouter, which lets developers route requests across models based on factors such as capability and price, had reportedly been valued at $1.3 billion after its May funding round.

CURSOR IS NOW A PART OF SPACEX

SpaceX has acquired Cursor to enhance AI model training using its extensive GPU resources. This acquisition enables Cursor to develop stronger and more cost-effective AI models. Grok 4.6, recently released, demonstrates the potential of this collaboration.

THE STATE OF OPEN MODELS IN 2026

Hugging Face reviewed major developments across the open-model ecosystem from January through August, using ecosystem data to highlight how model releases, tooling, and adoption had shifted since its spring report.

UNDERSTANDING AGENT MEMORY

This analysis compared persistent agent memory built from curated files, automatically maintained structured stores, and learned experience. It examined the approaches under controlled models and agentic benchmarks to show how different memory representations affected performance.

GLM-5.3: HOW CHINESE LABS KEEP STRIDE WITH THE FRONTIER

Z.ai has a time-to-release of likely days, rather than months as with OpenAI or Anthropic. While OpenAI and Anthropic likely have far better internal models, the companies tend to take months to release their models to the public. Z.ai probably cares a little more about public benchmarks than its US counterparts, but it is not benchmaxxing to the point where it is too obvious with GLM-5.3. The RL industry is taking off in China, very much driven by US data companies selling to Chinese model labs.

ON DWARKESH PATEL'S PODCAST WITH RYAN GREENBLATT

Dwarkesh Patel and Ryan Greenblatt's podcast debate focused on recursive self-improvement (RSI) and AI alignment challenges, highlighting their differing views on AI's capabilities and risks. Greenblatt argued for AI's efficiency in certain R&D tasks but acknowledged difficulties in verifying alignment, suggesting a complex training process focused on narrow tasks could lead to misaligned models.

MATHCODE

MathCode is a frontier mathematical coding agent. It has a built-in math formalization engine that converts problems from plain language into Lean 4 theorems and attempts formal proofs. It features a persistent Lean REPL, reusable theorem and axiom libraries, agent proving, and an Obsidian knowledge graph. The math formalization and proving pipeline is based on the AUTOLEAN project.

INTRODUCING CUSTOM AGENTS

Google has introduced Custom Agents in Antigravity 2.0 and the Antigravity CLI, with the Antigravity IDE following shortly. Custom Agents are specialized, file-based configurations that define a particular role with its own scoped instructions, tools, and constraints. The system keeps users' active contexts clean, minimizes token overhead, and gives them a predictable partner for specific tasks. Custom agents don't replace skills and dynamic subagents - they just provide even more customizability for another level of optimization.

LITTLELEARNER

LittleLearner is a language model that only knows what a 5th grader knows. The hosted 5B model can be run live in browser. LittleLearner was created as part of an experiment in creating controlled sandboxes for studying how models acquire knowledge.

MICROSOFT'S FULL-BANDWIDTH TRANSFORMERS

Full-bandwidth transformers fed the previous token's top-layer hidden state back into the model alongside the next token embedding. This allows latent computation to continue across decoding steps.

GTM ENGINEER, APPLIED AI AT TLDR ($175-205K BASE + $40-60K BONUS, FULLY REMOTE)

TLDR is hiring a GTM Engineer to join our Applied AI team and own our AI-native GTM stack. We're looking for someone comfortable building AI agents and working with HubSpot. Click here to learn more [22]!

OPENAI PAID $100 FOR A 4.2% CEREBRAS STAKE WEEKS BEFORE ULTRAFAST LAUNCH

Days before OpenAI previewed Ultrafast, a service tier that can run GPT-5.6 Sol at up to 750 output tokens a second, OpenAI exercised every vested Cerebras warrant share, acquiring 10,033,508 Class N shares at $0.00001 each for about $100 in cash. The stake has an implied value of near $2.3 billion, though the shares carry no votes. Cerebras is now running every competitive speed offering at OpenAI. OpenAI has yet to publish price, model ID, or general-availability date for Ultrafast.

DARIO AMODEI ON REGULATION AND THE MESSAGING AROUND AI

Either concentrating AI in the hands of a chosen few companies and politicians via regulation or distributing it widely is a false choice. Those in that frame of mind often underrate the decentralizing power of objective and fair institutional processes. The public's negative view of AI is fundamentally a crisis of trust. Ordinary people don't trust companies, governments, or the tech industry. AI companies have still yet to deliver on their big promises to benefit the world.

GOOGLE REPORTEDLY TAPS AMD TO DESIGN NEXT-GENERATION TPU — HYBRID AI ASIC COULD INTEGRATE ON-PACKAGE CPU CORES FOR REINFORCEMENT LEARNING

Google appears to be exploring a new kind of TPU that combines its proprietary accelerator technology with on-board general-purpose cores for CPU-heavy workloads.