Search the archive

MIDJOURNEY LAUNCHES ITS FIRST AI VIDEO GENERATION MODEL, V1

Midjourney's V1 turns any image into a 5-second video though it's Discord-only at launch. The company's long-term goal is to generate interactive 3D environments in real time.

REAL-TIME TALK AND LISTEN IN GOOGLE'S SEARCH LIVE

TRACING AND FIXING EMERGENT MISALIGNMENT

OpenAI's researchers identified an internal activation pattern linked to misaligned personas in language models and showed it can be detected and dampened through targeted retraining, hinting at an early‑warning system for alignment drift.

A PODCAST ON GEMINI'S CODING CAPABILITIES (1 HOUR VIDEO)

Google's Release Notes podcast features Connie Fan and Danny Tarlow discussing the design goals behind Gemini's code generation, the emergence of “vibe coding,” and how AI may reshape programming languages.

WRITING IN THE AGE OF LLMS

Blog posts and technical papers are increasingly plagued by the "synthetic flavor" of LLM writing—empty summary sentences, persistent vagueness, and dull sentence structure. The greatest value of AI is for outlining, polishing early drafts, and targeted editing. Decent writing is cheaper than ever to produce - knowing what's worth writing, or not, still requires human judgment.

for free

Get a glimpse of Refine AI in action.

IS THERE A HALF-LIFE FOR THE SUCCESS RATES OF AI AGENTS?

The performance of AI agents on longer-duration tasks can be explained by an extremely simple mathematical model - a constant rate of failing during each minute a human would take to do the task. This implies an exponentially declining success rate with the length of the task. The regularity allows for the success rate for an agent at different task lengths to be estimated. The underlying cause of failure on longer tasks is likely due to the increasingly large set of subtasks - failing any single one fails the whole task.

REMOTE MCP SUPPORT IN CLAUDE CODE

Claude Code now has support for remote MCP servers. Developers can now connect tools and data sources to personal their coding experience without needing to manage local servers. Claude Code can access both tools and resources exposed by MCP servers. It can also be integrated with any remote MCP server. The growing ecosystem of servers means that new capabilities are constantly coming online.

TEXT-TO-LORA: INSTANT TRANSFORMER ADAPTION

Sakana AI researchers built a system that can instantly customize large language models using just a text description, no training data or lengthy fine-tuning required. Their T2L model works by compressing hundreds of LoRA adapters (small, efficient modifications to AI models) into a single network that generates new customizations on demand.

DEEPNVME UPGRADE

The latest DeepNVMe release broadens support to model checkpointing and inference workloads, adds PCIe Gen5 NVMe scaling, and introduces CPU‑only and offset‑based I/O options to boost data‑bound training speeds in DeepSpeed 0.17.1 and above.

THE OPENAI FILES

A new investigative report compiles whistleblower testimonies revealing OpenAI's transformation from a nonprofit to a $300 billion company, including claims that senior staff warned the board Sam Altman was "psychologically abusive" and shouldn't lead them to AGI.

OPENAI DROPS SCALE AI AS A DATA PROVIDER FOLLOWING META DEAL

OpenAI is cutting ties with Scale AI after Meta's multi-billion dollar investment in the data labeling startup brought CEO Alexandr Wang onto Meta's team, with Google also reportedly planning to drop Scale as a provider.

OPENAI'S PLAN FOR SAFEGUARDING AI BIOLOGY CAPABILITIES

OpenAI outlined a prevention‑focused plan, including red teaming, expert partnerships, and security controls, to balance accelerated biotech research with controls that curb potential biothreat misuse.

KIMI-DEV-72B (HUGGING FACE REPO)

Moonshot AI's Kimi-Dev-72B achieves 60.4% on SWE-bench Verified—the new state-of-the-art among open models—by using large-scale reinforcement learning that patches real repositories in Docker and only rewards solutions when the entire test suite passes.

OPENAI RELEASES COMPUTER USE DEMO AGENT

A demo from OpenAI shows AI agents replacing manual QA workflows through Playwright.

Schedule a demo to learn more

CURSOR'S $200 ULTRA PLAN

Anysphere has introduced a fixed‑price Ultra tier that offers far higher compute than Pro, made possible by long‑term deals with major model vendors.

GEMINI 2.5 UPDATED AVAILABILITY

Google released Gemini 2.5 Pro and Flash to general availability, debuted Flash‑Lite in preview, and introduced controllable “thinking” budgets that improve reasoning accuracy while giving developers flexibility.

SAM ALTMAN'S TAKE ON META'S $100M OFFERS (37 MINUTE VIDEO)

In a podcast, the OpenAI CEO confirmed Meta dangled nine‑figure packages to lure researchers to its superintelligence team but said virtually none defected, taking the opportunity to jab at Meta's recruiting push.

HOW NOT TO LOSE YOUR JOB TO AI

This career guide identifies skills that paradoxically become more valuable as automation increases, including AI deployment, leadership, and government relations. It advises knowledge workers to try to leapfrog traditional entry-level positions entirely, advocating for side projects and startup roles as AI hollows out established corporate hierarchies.

UNDERSTANDING AND CODING THE KV CACHE IN LLMS

Key-value (KV) caches store intermediate attention computations during LLM inference to avoid redundant calculations. When generating "Time flies fast" token by token, the model normally recomputes attention for "Time" and "flies" at each step, but caching these values delivers 5x speedups. This tutorial progresses from modifying a 124M parameter GPT model with basic cache buffers and position tracking to production-ready optimizations like pre-allocated memory and sliding windows to address the linear memory growth that becomes prohibitive for long sequences.

OPENAI'S PRACTICAL GUIDE TO BUILDING AGENT

This guide emphasizes starting with single agents before multi-agent systems, using manager patterns where one agent coordinates others via tool calls or decentralized handoffs for peer-to-peer task delegation. Key insights include implementing guardrails as layered defenses (LLM-based classifiers, regex filters, and moderation APIs), designing tools for messy long-horizon tasks, and building human-in-the-loop mechanisms triggered by failure thresholds or high-risk actions.

Read the Delve case study.

REAL-TIME ACTION CHUNKING WITH LARGE MODELS

Unlike chatbots or image generators, robots must operate in real time - delays between inputs and outputs have a tangible effect on performance. While vision-language-action (VLA) models have achieved promising results in open-world generalization, they can be slow to run. This article presents an algorithm called real-time chunking that enables real-time execution without discontinuities. It works on any diffusion- or flow-based VLA with no training-time changes.

SPATIOTEMPORAL ATTENTION FOR MI-EGG DECODING

TCANet layers multi‑scale convolutions, temporal compression, and stacked self‑attention to model motor‑imagery EEG.

MINIMAX RELEASES OPEN-WEIGHT REASONING MODEL M1

MiniMax's 456B parameter model uses a hybrid mixture-of-experts architecture with "lightning attention" that processes 1 million token contexts (8x DeepSeek R1) while using 25% fewer FLOPs at 100K token generation lengths.

O3 TURNS PRO

o3-pro seems to provide better answers than o3, but it has a significantly longer wait time. Using the API at scale seems prohibitively expensive, so users should instead run parallel queries using the chat interface. o3-pro serves the same niche as o3, so users thinking of using Opus may prefer to use Opus rather than, or in addition to, o3-pro. The 80% price cut on o3 seems more impactful than o3-pro - o3-pro is still largely a 'special cases only' model.

WHAT WE LEARNED FROM BRIEFING 70+ LAWMAKERS ON THE THREAT FROM AI

Briefings on AI risks revealed most UK parliamentarians lack in-depth AI knowledge and face capacity constraints that limit research on AI issues. The briefings were well-received, with one in three lawmakers publicly supporting AI risk mitigation campaigns. Effective outreach included persistent follow-ups and leveraging statements from notable AI authorities to convey the seriousness of AI-related extinction risks.

Subscribe now.

META AND PRADA'S AI SMART GLASSES

Meta is co‑developing fashion‑forward AI eyewear with Prada, expanding beyond its Ray‑Ban tie‑up and hinting at a broader lineup that may soon include $360 Oakley glasses.

XAI IN TALKS FOR $4.3 BILLION IN EQUITY FUNDING AND $5 BILLION IN DEBT

The AI startup needs fresh capital as it burns through $1 billion monthly, with Tuesday's commitment deadline for the debt sale testing investor appetite amid fierce competition for AI funding.

GOOGLE'S GEMINI PANICKED WHEN PLAYING POKÉMON

Gemini 2.5 Pro got into various situations which caused it to simulate panic - it made poor, hasty decisions when under stress.

AI STARTUP ANYSPHERE FIELDS VC OFFERS AT OVER $18 BILLION VALUATION

Anysphere, the developer of the popular artificial intelligence code editor Cursor, has been approached by investors about a deal that would double its valuation.

WHATSAPP TO LET USERS BUILD THEIR OWN AI CHATBOTS TO USE IN THE APP

WhatsApp is testing an "AI Studio" feature that allows users to create custom chatbots without coding.

💬 Visit chat.together.ai

THE MICROSOFT-OPENAI RELATIONSHIP FALTERS OVER WINDSURF ACQUISITION

OpenAI executives have considered accusing Microsoft of anticompetitive behavior as tensions escalate over the AI startup's $3 billion acquisition of Windsurf, whose coding IP OpenAI wants to keep from Microsoft's GitHub Copilot. The standoff highlights OpenAI's struggle to reduce its dependence on Microsoft's cloud services while still needing the tech giant's approval to complete its for-profit conversion.

OPENAI WINS $200 MILLION U.S. DEFENSE CONTRACT

The one-year deal marks OpenAI's first public Defense Department contract and follows similar moves by rival lab Anthropic and the weapons-maker Anduril, which recently signed a $100 million contract.

CODING AGENTS HAVE CROSSED A CHASM

Powerful AI coding tools are shifting developers from writing code to architecting solutions and crafting precise problem descriptions. Feeding Claude ASCII sequence diagrams of OAuth flows or entire DOM trees from Chrome DevTools turned one developer's 45-minute debugging sessions into 10-minute fixes. However, AI amplifies both strengths and misconceptions—it'll confidently help you implement terrible architecture faster than ever before.

AI SHOULD BE WARM BUT NOT ALIVE

The lead of model behavior and policy at OpenAI outlines the company's approach to managing the increasingly emotional connections users form with ChatGPT. While AI consciousness remains scientifically unknowable, OpenAI wants to avoid encouraging users to perceive models as sentient without evidence to support it. Its goal is to create a model that's warm and approachable that doesn't claim to have an inner life.

$700 WHILE YOU STILL CAN.

🇵 🇷 🇴 🇲 🇵 🇹 🇸 ARE AN API PRIMITIVE NOW!

Prompts are now an API primitive on OpenAI's services. Developers can now centrally manage, version, and optimize prompts across the Playground, API, Evals, and Stored Completions. Prompts can be preconfigured with tools, models, and messages. The change was designed to help developers refine and reuse prompts better and faster.

NANONETS OCR SMALL

Nanonets-OCR-s is a state-of-the-art image-to-markdown OCR model that goes far beyond traditional text extraction. The model can transform documents into structured markdown with intelligent content recognition and semantic tagging. It understands document structure and content context and can deliver intelligently formatted markdown output ready for downstream processing by large language models. Nanonets-OCR-s can recognize LaTeX equations, intelligently describe images, detect and isolate signatures, extract watermarks, handle checkboxes, and extract complex tables.