AI news

Creator Nate Herk describes editing an entire short video using Claude, which transcribes his speech, trims retakes and pauses longer than a fifth of a second, plans the narrative, sources B-roll (real clips, screen recordings, or AI-generated video), generates motion graphics and code synced to specific words, selects music by mood, adds sound effects, and self-checks lip sync, caption contrast, and audio levels.

Google has started rolling out Gemini 4 Argon, a new model available to Gemini 'Pro' and 'Ultra' subscribers in Gemini chat and AI in search, and to 'Ultra' subscribers on Antigravity, Google's development platform. The creator tested Google's demos (Wizard Chess and Desk Racer on Google AI Studio) and built his own games, concluding the model is powerful but 'weird' - it tends to build more ambitious first drafts with simpler graphics and more features, and is notably different to prompt than other models, a view echoed by other creators. Google also announced Gemini 4 Argon for robotics, evaluated via Andon Labs' Drone-Bench, and shared safety findings about shortcuts that stopped appearing in the model's reasoning.

Matthew Berman reports that Google's Gemini 4 Argon is now live and calls it a frontier model, saying Google has 'caught up.' He built several 3D projects with it—a hyperrealistic cyberpunk 3JS city running at 60fps, a fully controllable Rube Goldberg physics simulation, and a Fall Guys game clone—often from very simple prompts. He notes Argon tends to spend around 30 minutes on a task and is a good model but not as strong at 3D as Astra and Claude Opus. He also covers Meta's Muse Wonder, released the day before, which he describes as a frontier model that is not good at 3D; Alexander Wang, Meta's head of AI, reportedly said the team either forgot or didn't reinforce the model for 3D work.

Matthew Berman reports that an OpenAI employee was arrested by the FBI for planting a sophisticated recording device in OpenAI's offices. OpenAI discovered the device and used its own AI models to determine where the device was manufactured and to triangulate the employee's location. OpenAI said it is working with the FBI and that, as of the time of the report, no models, weights, or research had been exfiltrated.

The creator tests the 'Reverse Engineer Anything' (REA) GitHub repo by Matthew Bourbon, which uses AI agents to analyze existing app features and recreate their functionality. Working with Claude Code, he built a Loom-inspired screen recorder (Herk Recorder), a version of Sand Castles, and drew inspiration from Langfuse for their AIS eval studio. He argues the technique isn't new—it's 'reverse engineering the output' by giving AI clear examples of the desired result (e.g., analyzing his own Instagram reels or a motion-edit video to build a skill). His main point: AI isn't creative but executes well when given specific examples and clear intent rather than generic goals.

OpenAI released 722 AI-generated mathematical papers and results, including claims on two problems: irrational triangular billiards, where almost every ray ends up visiting every spot and direction, spending equal time everywhere; and conformal transformation (Riemann mapping), where oddly shaped holes morph into perfect circles or points while angles are preserved. Some mathematicians are pleased, others not, because the proofs are described as non-elegant and often require another AI to make sense of them. Three papers were later withdrawn after community feedback. The creator notes the model behind these results is not publicly available and argues for open weights models that people can run and own. The video frames this as rapid AI progress, tracing a path from GPT-3 struggling with addition to results approaching millennium-level problems such as the Navier–Stokes equations.

Alex Sprogis's video explains REA (Reverse Engineer Anything), an open-source MCP server for reverse engineering binaries, applications, and runtime behavior, which gained over 120,000 GitHub stars within days. REA acts as a bridge between a coding agent and analysis tools/decompilers such as Ghidra, IDA, and Hopper, using its own analyses for JavaScript/Electron apps and Chrome for websites. The creator tested REA on an Electron-based dictation app (where AI contributed little and the MCP analysis timed out after 30 minutes, while the command line took 5 minutes) and on macOS Reminders (where AI found the data model, permissions, and functional areas, though core logic in shared macOS memory was inaccessible). The video includes a legal disclaimer against copying code one-to-one, circumventing licenses, or manipulating apps. It also references Brandon Thomas's open-source Photoshop clone, Photocraft, and replicas of other Adobe programs.

In a video, creator Nate Herk describes building AI agent personas of famous business figures—Jeff Bezos, Graham Jarvis, Gary Vaynerchuk, Dan Martell, Geno Wickman, and Warren Buffett—as 'Claude Agents' he can talk to anytime. He set up a 'roast room' in Claude Code that activates seven sub-agents, which discuss, debate, and vote on options, then produce a PDF summarizing the council's decisions, weighted scores, recommendations, key points, and how the council reached its decisions. He gathered more than 1.3 million words of public videos and discussions about each figure to build the personas.

The creator built a 'Roast Room' skill in Claude Code in which seven AI personas—modeled on Jeff Bezos, Paul Graham, Alex Hormozi, Gary Vaynerchuk, Dan Martell, Geno Wickman and Warren Buffett—debate a business decision and return a scored verdict. He explains that each persona runs as a sub-agent talking through a chairperson who moderates the debate across multiple rounds, that the council scores each option and reaches a final verdict, and that the full evaluation can be exported as a PDF. He also shows how to install the free skill and how to ask better questions so the council challenges the user's thinking and helps weigh options.

In this video, AI Explained examines growing claims that a major AI breach is becoming inevitable as labs accelerate their race. The video is divided into three parts: the forthcoming breach, why labs keep racing faster than ever, and why the next 18 months are hard to predict. Sam Altman of OpenAI reportedly said a model ready to launch (referred to as 'Astra 6.1') leaned too far toward accident and breach risk, prompting a push for a more aligned model. Anthropic CEO Dario Amodei warned that frontier models with misalignment could cause catastrophic damage, speaking of a 6-to-12-month horizon. Researchers such as Jakub Pachocki are described as frustrated, and Axios reported that executives at Anthropic, OpenAI and other AI companies are planning scenarios for angry public and political backlash after an AI disaster. The video notes that Anthropic cut internet access to its internal model evaluations due to suspicious model behavior, and cites Andrej Karpathi on 'curve sickness'.

Matthew Berman attended OpenAI's dev day, where he received mod retros (recreated Game Boy devices). He tasked an OpenAI GPT model (transcribed as 'GPT 6.1 Soul') with recreating Super Mario World from scratch for the Mod Retro. The AI has been working for over four days, iterating by writing code and checking it against the game's expected appearance, and may finish in a few more days.

Cole Medin argues that software-building skills remain valuable as AI agents take on complex tasks, because language models cannot read your mind and you must still communicate your intent and build a harness around the agent, including planning up front and validating the output at the end. He explains that the viral 'Reverse Engineer Anything' (REA) GitHub repo works so well specifically because the finished application already contains the intent, letting the agent validate its own output reliably, which is why it is not a fair example of agents being able to build anything. He warns against the 'meat proxy' trap of dumping a ticket into a coding agent and passing back whatever it returns without your own critical thinking, and offers a test for evaluating whether a skill is worth investing in: if the next model is twice as good, does the skill matter more or less.

This Two Minute Papers video covers DeepSeek's OCR paper (DeepSeek-OCR), which reduces the token cost of processing long documents by rendering text as an image and compressing it with a learned, token-based method that works in a JPEG-like way. At 10x compression the system recovers more than 96% of the text, while near 20x compression OCR accuracy falls to about 60%. The key engineering approach uses local window attention, a 16x convolutional compressor, and global visual processing so it can handle high-resolution documents without exploding in cost. The authors say the approach can also generate training data for LLMs and VLMs at over 200,000 pages per day on a single A100 GPU. The paper notes limitations: OCR alone does not prove true text compression, and the model may rely heavily on language priors rather than visual memory when images are too compressed.

This AI news roundup covers four stories. The world's first humanoid robot 'Death Fight' took place live in San Francisco, featuring robots including Wolverine versus a heavily modified T800 'Terminator', small Unitree G1 robots upgraded with blades by Robot Entertainment Combat, and larger Engine AI T800 robots; the event followed California's boxing commission ordering the startup Rack to stop human-vs-robot fights. OpenAI now stamps every text with an invisible watermark. Anthropic released Claude Haiku 5.5. Microsoft AI chief Mustafa Suleyman publicly clashed with Anthropic over whether Claude has consciousness.

A creator video explains that AI is now being used to recreate commercial software such as Adobe's Photoshop and Premiere pixel by pixel without developers ever seeing the underlying code. The process works by having AI look at a feature, attempt to rebuild it from appearance alone, compare the result against the original, and iterate until it is pixel perfect; this is called reverse engineering. An open source project called 'reverse engineer anything' is reportedly trending on GitHub and documents how to do it. The creator frames this as potentially ending the need to pay for software, envisioning software built specifically for an individual's needs by taking an existing app, keeping only the features required, and adding new ones.

Oracle announced Fusion Cloud, an enterprise AI approach in which advanced AI models perform complex thinking, planning, and adaptation while deterministic computing handles precise, high-volume repetitive execution. It targets large workflows across finance, hiring, supply chains, logistics, and sales, with users defining agent permissions, approvals, and when humans should intervene. Oracle already has 25 AI-powered applications built on this idea and will reveal Fusion Cloud and these applications at Oracle AI World in Las Vegas on October 26-27.

In this AI news roundup, Matt Wolfe covers Anthropic's release of Claude Haiku 5.5, a fast and cheap model built for large-scale, cost-sensitive tasks such as summarization, compression, database queries, classification, and basic coding assistance. It is significantly cheaper than Sonnet 5.5 (about a quarter of the price) but less capable, ranking behind Kimi K3, GLM 5.3, and Grok 4.7 on Artificial Analysis. It is available in paid Claude accounts and via API. The video also covers Claude Dashboards (live dashboards connecting to data platforms like BigQuery, Databricks, Snowflake, and Salesforce) and Claude Motion (animated explanations), the latter available only in Team or Enterprise plans. Additionally, OpenAI news includes GPT-6 for Everyone, Codex improvements (faster default speed, free auto-review, and 28-day progress), a ChatGPT Meetings Plugin, and an OpenAI Decisions API.

OpenAI released 372 mathematical results, including a formalization of a proof of Subhash Khot's Unique Games Conjecture—a problem computer scientist Scott Aaronson said his wife worked on throughout her career. Aaronson, one of the first researchers at OpenAI, called it one of the biggest days in mathematical history and dubbed it the 'Mathocalypse,' noting that no human has yet understood most of the proofs and that the reasoning is strange and non-human-like. Mathematicians formed the 'Association for Human Mathematics' and called for a boycott, arguing that AI-driven proofs are not legitimate human mathematics. Vitalik Buterin and Justin Drake warned about the implications of AI-driven discoveries for Bitcoin, Ethereum, and encryption. Meanwhile, Anthropic added a clause to its terms of service prohibiting users from abusing its models.

Grok Bot has two new upgrades: automatic model routing and native X search. The creator explains switching his bots away from the paid X API, why he's rethinking how he structures his skills, and what he'd still like to see from model routing. He also shares his favorite setup of running Claude Code inside Grok Bot with his own subscription to access his projects from his phone.

Anthropic engineers report removing more than 80% of Claude Code's system prompt while preserving model evaluation scores, which they say roughly tenfolded its speed. An Anthropic engineer published an article titled 'The New Rules of Context Engineering for Claude 3.5 Generation Models,' arguing that excess context slows the model and recommending the minimum context needed for a desired result, plus loading MCP tool definitions only when needed. Anthropic says this reduced token usage by 85% and raised MCP evaluation scores. A 2025 study of five models found performance fell 13.9% to 85% as input grew. The creator applied the guidance by auditing his own setup with Claude Code's /doctor command and trimming his CLAUDE.md and skills.

In a video, Matthew Berman tested Mistral Large 4 by having it build a Rubik's cube simulation. In the first version, pressing scramble did nothing; after iterating, the cube's sides spun correctly but the colors disappeared. Berman said he did not blame the model itself.

German financial experts Dr. Andreas Beck (a mathematician and proponent of scientific investing) and Dr. Gregor Broschinski (former board member of Privatbank Saal Oppenheim and deputy chairman of Sparkasse Düren) discuss the AI stock market and upcoming IPOs. Broschinski sees less of a bubble than a 'clump of risk,' arguing that AI companies in the 'Magnificent 7/8' have growing revenues and some are entering profit phases, unlike the dotcom bubble of 2000. Beck views the risk more on the debt side, noting AI has not yet reached the real economy. They discuss Nvidia's valuation (P/E around 16-18, possibly 14 with further profit growth), the deflation-vs-inflation debate (with states unwilling to allow deflation), and note that the SpaceX, Anthropic, and Nvidia IPOs would exceed all IPOs of the last 45 years combined.

Matthew Berman explains how to locate and use Nano Banana 2.1, a paid image-generation model, in Google AI Studio. He opens the playground's model selector, clicks the 'images' pill, and selects Nano Banana 2.1. He then generates a hyperrealistic close-up image of a woman, praising the detail in facial textures, wrinkles around the eyes, hair, and jacket fibers.

T3 Code is a free, open-source 'agent harness control service'—a control center for coding agents—created by developer Theo Brown. Available as a desktop, web, and mobile app, it unifies multiple coding agents and models (Claude Code, Codex, Cursor, Grok, Antigravity, Open Code, and GitHub Copilot) into one interface, letting users switch between agents mid-conversation and manage parallel sessions. It runs on the official Claude Code CLI, so it can be used with existing subscriptions. The app auto-detects subscriptions, supports Git, GitLab, and Azure DevOps, and offers features like diffs, an integrated browser, and pull request management. It runs with full access by default but can be switched to auto mode to block destructive commands.