Mistral's Le Chonk, PhotoCraft, and Robot Chonk: 6 videos from October 5–11

The short version

Mistral is BACK! (Le Chonk)

• Mistral opened a public preview of Mistral Large 4 on October 6 and promises the weights by the end of the month. The company calls the model ML4, and officially calls it le Chonk. 1
• ML4 is a 1-trillion-parameter natively multimodal mixture of experts with 52 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. It costs $1.36 per million input tokens and $4.18 per million output tokens. 1
• The coding numbers Mistral publishes are 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.0, both agentic coding evaluations, with a combined Coding Agent Index of 49.8%. On AutomationBench, 657 business workflows across Gmail, Sheets, Slack and Salesforce, ML4 scores 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro 0813. 1
• Cybersecurity is the distinctive section. Mistral reports a top-five finish on Artificial Analysis's independent Cyber Index, 82% on one of its tests that asks a model to reproduce a real open-source vulnerability and then patch it, and 93% on Cybench's 40 competition exercises. Until the weights ship, cybersecurity leaders, vetted partners and state authorities get the same model with reduced moderation and expanded cyber capabilities. 1
• Berman's counterweight is the harness problem. The preview arrives as an API only, so trying it means getting an API key, and he could not get it working well inside a third-party agent harness: with OpenCode the session ended because the model's thinking tokens filled the context window and it stopped returning anything. He names Cursor as the exception he has found, and puts the closed US frontier about six months ahead. 2

Everything has been solved

• Decompilation means recovering readable source code from a compiled game with no access to the original. Berman's example of the old pace is Super Smash Bros. Melee, which a community has been decompiling for years. His example of the new pace is Chris Lewis, who spent nearly two years decompiling Snowboard Kids by hand and then did the same job in 84 days with AI. 3
• The reason AI suits the work is the loop, not the code volume. Each attempt can be checked against how the game actually behaves, so many agents can grind in parallel against a verifiable target. Berman's own run is GPT-6.1 Sol rebuilding Super Mario World for the Mod Retro handheld given out at OpenAI's Dev Day, which had been working for more than four days when he recorded. 3
• His Photoshop evidence is PhotoCraft, one of ArtCraft's open-source Rust applications. Its project pages describe it as an early-alpha, clean-room reimplementation of Photoshop built from public specifications and observed behaviour, released under MIT or Apache-2.0. It implements layers, masks, adjustment layers and layer styles, opens and saves layered PSD and PSB documents, and round-trips 134 of 135 real test files byte for byte. 45
• Game mashups, as Berman charts them, come in three kinds: swapping a character's assets so the game behaves as before, running both games and synchronising events across them, and porting a mechanic like skate physics into a foreign engine. Only the third needs the original behaviour understood well enough to rebuild it, which is the part he says AI unlocked. 3
• The wider claim rests on mathematics. OpenAI's October 6 release published a broad set of new results on open problems, all produced by an unnamed internal frontier model and shipped on GitHub with Lean formalisations of many proofs. Berman reads the pattern as any problem with a checkable answer being reachable, and ends on the consequence for software businesses: if the code itself was the moat, he says, it is gone, and taste and support are what remain. 36

12 Grok Bot Use Cases That Feel Illegal

• Grok Bot, from SpaceXAI, is a set of always-on agents that each keep their own cloud computer, sign into apps and websites that have no clean API, and come back to the user only when something needs approval. It launched in beta on August 11 for the company's own subscription tiers, with usage counted separately from Grok and Cursor plans. 7
• The money-saving bot is the most reusable. Told to inventory subscriptions across three Gmail accounts and negotiate, it found more than $1,000 a month in fixed costs and drafted the renegotiation and cancellation emails for him to send. The same bot files claims for unclaimed property on a state website. 8
• Two examples run through real accounts rather than a chat window. His car bot was connected to a Tesla Model Y through Tesla's official Fleet API, so he can ask for the climate and seat heaters to come on and for a destination to be set. His family bot reads school and sports emails, posts one-line summaries into a family Telegram group, archives the originals and writes each event into a shared calendar with the other adults invited. 8
• The back-office set covers meeting notes, messages and disks. A meeting bot pulls transcripts from his recorder, writes a summary with owners and dates, and adds action items to a to-do app; a lead bot triages direct messages across Instagram, LinkedIn and X and drafts replies to sponsors; a weekly cleanup bot found about 50 GB on his Mac Studio it judged safe to clear, and its default is to list the files and wait for his go-ahead. 8
• Every use case ships as a Grok Bot template plus the underlying prompt, so the prompts can be pasted into another assistant. Berman's own rule is repeated across them: the bot proposes, and a human presses send. 8

Robot Chonk Cleans the Floor

• The milestone is a first: the robot picked a toy off the floor, noticed when it dropped the toy, found it again and put it in a bin. Neither the object nor the bin carried a marker, both had moved since the last attempt, and he trained nothing extra for the run. The clips run at 10× speed because the real motion is slower. 9
• The robot is Sourccey, from Vulcan Robotics, described by its maker as an open-source developer robot for real-world data acquisition and physical AI. Its hardware, electronics and software are published, it follows the LeRobot toolchain, and the company's roadmap has the platform open-sourcing in October 2026 with robots shipping in November. 10
• The hardware choices answer the task. Cameras sit in the grippers, so the robot can see what it is holding; the arms slide up and down the body, which gives the reach of a humanoid without the balance problem; five cameras and a 2D lidar cover the room. His one complaint is that the arms have five degrees of freedom, not six. 9
• DeepSeek V4.1 Flash does the whole job: planning, motion and control, driven by a numbered prompt and the camera images, with the policy itself coming from the general model. GLM 5.3 Flash also works, and he notes that the speed of the model on his new workstation is what made the experiment practical. 9
• He explains why he stopped reaching for decision models. When the task goes wrong — a dropped toy, a dog walking into frame — a generalist multimodal model can look at the picture and recover, while the decision models he tried did not handle those failure modes. 9

The Billion Dollar AI Advantage Is Disappearing

• The demonstration is a code port. Sonnet 5.5 took a physics research codebase for rigid-body simulation, which has to keep collisions, friction and joints stable without the numbers exploding, and rebuilt it as a single HTML file that reproduces the original experiments. The same session also produced ray-traced and volumetric caustics. 11
• The prices behind his argument are Anthropic's own. Sonnet 5.5 launched on September 28 at $2 per million input tokens, $10 per million output tokens and $0.20 per million cache-read tokens, runs more than 30% faster than Sonnet 5, and costs up to 30% less per task. 12
• His price ladder puts the change in one line: Fable, then Opus 5.5 at roughly two and a half times cheaper with comparable quality, then Sonnet 5.5 at roughly five times cheaper than Fable. Cheaper, he reasons, means smaller. 11
• The argument is that comparable intelligence keeps arriving in smaller systems, so how well the parameters are trained matters more than how many there are. The prediction that follows is frontier-class models running on a laptop, and later on a phone. 11
• He labels that last part as his own reading, and says plainly that he could be wrong. No benchmark or vendor claim is offered for the laptop forecast; the evidence in the video stops at the code port and the price list. 11

Sparrow Studio Tutorial - Ep 1. - What is Sparrow Studio?

• Sparrow Studio is a desktop application for uploading, processing and managing ecology data in one place. Microsoft's own documentation describes it as the AI for Good Lab's unified desktop app, built on the PyTorch-Wildlife framework, and one route to running MegaDetector and species classifiers without writing Python. 13
• The inputs are broad. It takes historical archive data as well as near-real-time images, audio and video from 3G and 4G cameras, from Sparrow Edge field devices, or from third-party drones, and it integrates third-party sensors such as GPS tags. 14
• Analysis runs through a model zoo that the video puts at 53 models: 22 detectors, 27 classifiers and two aerial models. Each project is assigned a model or a set of them. 14
• The point of the platform is the review step. Predictions arrive attached to the data, misclassifications can be re-annotated by hand, and the corrected set can be exported as a PDF report or as CSV for other scientific tools. 14
• This episode is an introduction only. Creating a project, choosing its type and assigning models is left to episode two. 14