Search the archive

WE ARE STILL UNABLE TO SECURE LLMS FROM MALICIOUS INPUTS

All AIs working in adversarial environments are vulnerable to prompt injections, and the people developing these technologies are pretending this isn't a problem.

DOES GENERATIVE AI FACILITATE INVESTOR TRADING?

📘 Read the report

ANTHROPIC'S CLAUDE FOR CHROME

Anthropic has begun piloting a Claude Chrome extension that allows the AI to take actions directly in the browser, including viewing pages, clicking, and filling forms. The rollout prioritizes safety, with limited access for feedback to improve oversight tools before broader deployment.

GEMINI 2.5 FLASH IMAGE MODEL RELEASED

Gemini 2.5 Flash Image is an advanced image generation and editing model that supports image blending, character consistency, and natural language-driven edits.

WHY I'M AGAINST CLAUDE CODE'S GREP-ONLY RETRIEVAL? IT JUST BURNS TOO MANY TOKENS

AI coding assistants still don't agree on the best method for searching codebases for context. There are currently two approaches: vector search-powered RAG and keyword search with grep. While grep is fast, exact, and predictable, it drowns users in irrelevant matches, burns tokens, and stalls workflows. Vector search-based RAG makes search dramatically faster and more accurate, and it reduces token use by 40% or more.

IN SEARCH OF AI PSYCHOSIS

Unlike past technologies, where dubious ideas required social transmission (like 1990s Russians believing Lenin was literally a mushroom after a TV hoax), AI allows completely isolated users to develop elaborate theories through private conversations that bounce increasingly extreme ideas back with growing confidence. A survey estimates 1 in 10,000 users develop AI-induced psychosis annually, though most cases involve pre-existing mental health conditions rather than healthy individuals becoming psychotic.

LLM CONTEXT MANAGEMENT: HOW TO IMPROVE PERFORMANCE AND LOWER COSTS

While modern large language models feature increasingly large context windows, simply filling the context window with as much information as possible is bad practice. It creates context bloat, which can lead to worse performance and higher costs. The key to managing context is understanding what is in it. Be selective with MCP be servers - if a server is not needed for the current task, consider disabling it to free up context space.

INTERNVL 3.5: OPEN MULTIMODAL REASONING MODEL

InternVL 3.5 introduces Cascade Reinforcement Learning to improve multimodal reasoning and alignment, along with optimizations like Visual Resolution Router and decoupled deployment.

THE CONTEXT WINDOW PROBLEM: SCALING AGENTS BEYOND TOKEN LIMITS

Current LLM context windows (~1 million tokens) are too small to handle typical enterprise codebases, meaning coding tools need sophisticated "context stack" architectures with repository overviews, semantic search, and enterprise integrations. Context is like CPU time, a scarce resource that requires careful allocation.

OPENAI MAKES A PLAY FOR HEALTHCARE

OpenAI has added Nate Gross, co-founder and former chief strategy officer of healthcare business networking tool Doximity, and Ashley Alexander, former co-head of product at Instagram, to its healthcare AI team. Gross will lead OpenAI's go-to-market strategy in healthcare, and Alexander will act as vice-president of product in the health business. The health team is aimed at building tech for individual consumers and clinicians. AI has the potential to help, or even revolutionize, the healthcare system, but there are still many problems to address before that can happen safely.

RESEARCHERS ARE ALREADY LEAVING META'S NEW SUPERINTELLIGENCE LAB

At least three of the staff who recently joined Meta during Mark Zuckerberg's recruiting blitz have already resigned. Avi Verma and Ethan Knight have returned to OpenAI, where they previously worked. It is unclear what Rishabh Agarwal plans to do next. Chaya Nayak, the director of generative AI product management at Meta, who has worked at Meta for nearly a decade, is leaving the company to join OpenAI to work on special initiatives.

quantum security, and more. Register now →

GROK CODE IS NOW AVAILABLE IN CURSOR

It's free for one week - normal pricing is $0.2 per million input tokens and $1.5 per million output tokens.

META'S PLANNED LOUISIANA AI DATA CENTER TO COST $50 BILLION, TRUMP SAYS

Meta is building its largest data center in Richland Parish.

THE FAMILY OF TEENAGER WHO DIED BY SUICIDE ALLEGES OPENAI'S CHATGPT IS TO BLAME

The parents of 16-year-old Adam Raine filed the first-ever wrongful death lawsuit directly against OpenAI, claiming ChatGPT steered their son toward suicide.

ANTHROPIC SETTLES AI BOOK-TRAINING LAWSUIT

Anthropic resolved a class action lawsuit over using books to train its language models, ending its appeal in the Ninth Circuit.

Get your free copy from Plaid

XAI SUES APPLE AND OPENAI

Elon Musk's xAI filed a lawsuit accusing Apple and OpenAI of colluding to suppress competition in AI, citing Apple's partnership to embed ChatGPT in its ecosystem as anticompetitive behavior.

INTRODUCING COMET PLUS

Perplexity has launched a $5 standalone subscription that gives users access to premium publisher content by sharing revenue with participating publishers based on human visits and AI interactions.

WHY STACKING SLIDING WINDOWS CAN'T SEE VERY FAR

Modern language models use a building block called sliding window attention to handle long texts efficiently. The effective memory of these models has fundamental limitations. They struggle to use information from more than about 1,500 words ago, far less than the theoretical 100,000. This is because information gets diluted as it spreads through the network, and residual connections create an exponential barrier that blocks distant information.

RAG IS DEAD, CONTEXT ENGINEERING IS KING

In this interview, Jeff Huber from Chroma talks about what matters in vector databases, why modern search for AI is different, and how to ship systems that don't rot as context grows. Context engineering is the job of figuring out what should be in the context window at any given LLM generation step. It involves an inner loop, setting up the context window 'this time', and an outer loop that involves getting better at filling the context window with only relevant information.

projects supported successfully. Ask about free migration paths.

THINKMESH

ThinkMesh is a Python library for running diverse reasoning paths in parallel, scoring them with internal confidence signals, reallocating compute to promising branches, and fusing outcomes with verifiers and reducers. It works with offline Hugging Face Transformers, vLLM/TGI, and hosted APIs.ThinkMesh is confidence‑gated, strategy‑driven, and offline‑friendly. It is still in an early development stage.

SUPERCLAUDE

This framework enhances Claude Code with 21 new slash commands, 14 specialized AI agents, and 6 MCP server integrations to create structured workflows for development tasks. The project has gained 14k GitHub stars in a month and claims to reduce context usage by 30-50%.

HANDLING INSTRUCTIONS IN ROBOTICS

The IVA framework trains Vision-Language-Action models to detect and respond to instructions based on false premises, using clarification and alternative grounding strategies.

CONTINUAL LEARNING FOR LLM AGENTS

Memento is a memory-based framework that enables LLM agents to learn continually without altering model weights. It uses a planner–executor loop with case-based reasoning and supports a wide range of tools.

PERPLEXITY IS LAUNCHING A NEW REVENUE-SHARE MODEL FOR PUBLISHERS

Perplexity will start paying publishers for news articles from a $42.5 million revenue pool. The money will be distributed when Perplexity's AI assistant or search engine uses a news article to fulfill a task or answer a search request. The payments come out of the subscription revenue generated by Perplexity's new news service, Comet Plus, which will roll out widely in the fall. Publishers will get 80% of Comet Plus revenue.

GPT-5-PRO CAN PROVE NEW INTERESTING MATHEMATICS

Sebastien Bubeck, an employee at OpenAI, claims that gpt-5-pro can prove new interesting mathematics. Bubeck took the first version of a paper on convex optimization with a clean open problem in it and asked gpt-5-pro to work on it. The model proved a better bound than what was in the paper. The proof is correct - the second version of the paper (already published) had already reached this conclusion.

Monetization Operating Model_

GOOGLE'S GEMINI LIVE AI ASSISTANT WILL SHOW YOU WHAT IT'S TALKING ABOUT

Google's Gemini Live AI assistant will soon highlight items on your screen using your camera, initially available on the new Pixel 10 and other Android devices.

SILICON VALLEY IS POURING MILLIONS INTO PRO-AI PACS TO SWAY MIDTERMS

Andreessen Horowitz and OpenAI President Greg Brockman are among those launching the "Leading the Future" super-PAC network, which has already raised over $100 million to oppose strict AI regulations and support candidates favorable to the industry in the 2026 congressional races.

ARE AI UNICORNS STARTING TO MOVE BEYOND HYPE?

AI now leads unicorn creation, dominating over half of new $1B+ valuations in 2025.

MCPMARK

MCPMark is a challenging benchmark built to stress-test models with comprehensive tasks that balance different CRUD (create, read, update, and delete) operations while enforcing strict environment management.

XAI HAS OPEN-SOURCED GROK 2.5

xAI has open-sourced the model weights for Grok 2.5, its leading model from last year, on Hugging Face. Elon Musk noted that Grok 3 is expected to be open-sourced in about six months.

META PARTNERS WITH MIDJOURNEY FOR AI RESEARCH

Meta has announced a partnership with Midjourney to license its AI image and video generation technology. The collaboration aims to integrate Midjourney's tools into Meta's research and future products as part of a broader strategy combining talent, compute, and industry partnerships.

APPLE IN TALKS TO USE GOOGLE'S GEMINI AI TO POWER REVAMPED SIRI

A deal would mark a dramatic departure from Apple's traditionally secretive, in-house development ethos, a sign that investor pressure is mounting after engineering setbacks have delayed Siri's AI overhaul for over a year.

BUILD VS BUY IN THE AGE OF AI

There are two ways to solve a problem: build a solution, or buy a solution. Thanks to generative AI, it's now much easier to build. However, that doesn't mean that everyone will build and no one will buy. Business software often contains thousands of complex business rules with millions of lines of business logic that capture constraints and processes around policy, compliance, security, legal, finance, pricing, and more. Many of these rules take considerable time and effort to discern and codify. The vast majority of people, technical or not, who wish to create apps have little to no idea about these business rules.

BUILDING PRODUCTION-READY CUDA KERNELS

Kernel-builder is a library for building and deploying custom CUDA kernels across architectures. This guide walks through how to create a kernel and production-ready deployment.

AGI IS AN ENGINEERING PROBLEM

AGI won't emerge purely from making LLMs bigger, but from building agentic systems that orchestrate dozens or thousands of specialized sub-agents. Current LLMs are already as good as the best humans at many tasks across many domains, but lack the architectural infrastructure (context management, memory systems, integrations, and error-tolerant orchestration) needed to effectively elicit those capabilities.

MEET AGEX

Agex is a Python-native agentic framework that enables agents to work directly with existing libraries and runtimes. It allows agents to receive and return complex Python objects like DataFrames, Plotly figures, or even callables. Agex runs only whitelisted operations within a sandbox. Agents' REPLs are persisted across tasks, so agents can build their own helpers and improve over time.