Meta Re-engineers Engineers; Mythos Reveals 10k Security Holes; Robotaxis < Puddles

Today's AI Outlook: ⛅️

Meta’s AI Reorg Turns Engineers Into Data Labelers

Meta’s latest AI restructuring is getting messy... Again. According to AI Secret, engineering manager Sam Voigt survived the company’s latest 8,000-person layoff, then was "demoted" from manager to individual contributor. Other engineers reportedly got moved from infrastructure work to AI training data labeling.

Why it matters

The reporting frames Meta’s move as a pressure valve for headcount and AI development at the same time. Engineers who stay may see their roles flattened into model-support work. Engineers who leave face a colder tech job market, potentially without severance. That is a brutal corporate choose-your-own-adventure.

The Deets

• Meta reportedly shifted some senior infrastructure engineers into labeling AI training data.
• Manager spans are said to be moving from about one-to-eight to one-to-50.
• Meta already owns a major stake in Scale AI, a data-labeling company.
• AI Secret characterizes the shift as “engineered attrition,” designed to push some employees out without direct layoffs.

Key takeaway

Meta’s AI push is not just changing products. It is changing what technical careers inside the company look like.

🧩 Jargon Buster - Data labeling: The process of tagging or organizing examples so AI models can learn from them.

DeepSeek Cuts Prices, Undercuts US AI

DeepSeek permanently cut V4-Pro API pricing by 75%, putting pressure on U.S. frontier labs that still charge premium rates. The Rundown AI reported pricing of $0.435 per million input tokens and $0.87 per million output tokens, while AI Secret emphasized that the cut lands far below many U.S. rivals.

Why it matters

Cheap, capable models change the developer math fast. If DeepSeek can offer frontier-ish performance at near-commodity pricing, the battle moves from who has the flashiest benchmark to who can win everyday usage at scale. The pricing change also complicates the idea that export controls alone can slow China’s AI progress.

The Deets

• DeepSeek made a 75% price cut on V4-Pro permanent.
• The Rundown AI reported prices at $0.435 per million input tokens and $0.87 per million output tokens.
• AI Secret said Huawei’s Ascend 950 supernodes helped ease compute constraints.
• DeepSeek’s strategy appears built around good-enough performance at very low cost.

Key takeaway

The frontier model race is becoming a margin fight, and DeepSeek just brought a coupon cannon.

🧩 Jargon Buster - Token: A small chunk of text an AI model reads or writes, often part of a word.

Claude Mythos Finds 10,000 Security Nightmares

Anthropic’s Project Glasswing reported that Claude Mythos Preview and roughly 50 partners found more than 10,000 high- or critical-severity vulnerabilities in one month. The system is still gated because Anthropic says safeguards are not strong enough yet to prevent misuse.

Why it matters

AI security tools could help defenders find and patch vulnerabilities at scale, but the same capabilities could also help attackers. That makes Mythos a preview of the next cyber arms race: faster discovery, faster exploitation, faster patching, and a lot more caffeine for security teams.

The Deets

• Cloudflare reportedly found 2,000 bugs with a false positive rate better than human testers.
• Mozilla found and fixed 271 vulnerabilities in Firefox 150.
• Mythos flagged 6,202 high or critical issues across more than 1,000 open-source projects.
• Independent triage found that 62%, or nearly 3,900, held up.
• One partner bank reportedly used Mythos to detect and block a $1.5M fraudulent wire transfer.

Key takeaway

AI could become a major security advantage, provided defenders can patch faster than attackers can copy the playbook.

🧩 Jargon Buster - False positive: When a security system flags a problem that turns out not to be real.

🛠️ Tools & Products

Your AI Secretary Is Ready To Judge Your Calendar

The Rundown AI laid out a workflow for building an AI taskmaster using Codex or Claude Code. The assistant checks Slack, Gmail and your calendar each morning, then turns the chaos into a prioritized to-do list inside MonoNote.md.

Why it matters

This is where agentic AI starts to feel practical. Instead of asking a chatbot to summarize your inbox, you can set up a system that reviews yesterday’s feedback, rolls over unfinished tasks, updates rules and builds a daily work plan.

The Deets

• The workflow creates MonoNote.md and task-rules.md.
• Tasks are grouped by priority, source links and status checkboxes.
• The automation can learn from daily feedback.
• A weekly audit skill can scan recurring tasks and suggest what to automate.

Key takeaway

The most useful AI agent may be the one that quietly turns your digital debris field into a sane morning plan.

🧩 Jargon Buster - .MD file: A plain text file that uses simple symbols to format text. Think of it as a middle ground between a regular .txt file (no formatting) and a Word document (lots of formatting).

Robotaxis Still Have A Puddle Problem

Waymo reportedly paused service in Atlanta, San Antonio, Dallas and Houston after a robotaxi drove into a flooded Atlanta intersection and sat stuck for an hour. The pause followed a prior NHTSA recall related to flood-handling.

Why it matters

Autonomous driving has spent years solving maps, sensors, planning and edge cases. Water remains a beautifully low-tech villain. If robotaxis cannot handle flooding reliably, service expansion in storm-prone cities gets a lot harder.

The Deets

• A Waymo vehicle reportedly became stuck in a flooded Atlanta intersection.
• Service was paused in four cities.
• The issue followed a recall over flood-handling.
• AI Secret noted that both camera-first and sensor-heavy approaches have struggled with water-related edge cases.

Key takeaway

The next autonomous driving breakthrough may be less “general intelligence” and more “don’t drive into that.”

🧩 Jargon Buster - Edge case: A rare or unusual situation that can break a system even if it performs well most of the time.

Starship’s AI Infrastructure Dream Gets Partial Win

SpaceX’s Starship V3 flight reportedly hit a split outcome: Ship deployed 20 mock Starlinks, reached Mach 7 and splashed down as planned, while the Super Heavy booster lost its aft section during relight and was destroyed before controlled descent.

Why it matters

Reports tie Starship’s economics to the broader AI infrastructure race. Cheap, reusable heavy launch could help expand satellite connectivity, support AI inference globally and eventually enable space-based compute. Payload delivery worked. Reuse economics still need proof.

The Deets

• Starship V3 launched from Starbase.
• Ship 39 deployed 20 mock Starlinks and two heat-shield sensors.
• The Super Heavy booster failed before controlled descent.
• AI Secret notes Artemis 2028 still depends on a modified Starship for a human moon landing.

Key takeaway

Starship showed it can move payload. The bigger test is whether it can make the cost structure come back down to Earth.

🧩 Jargon Buster - Booster reuse: Recovering and flying the rocket’s first stage again to lower launch costs.

💰 Funding & Startups

Google Wants Founders To Build Agents That Do More Than Chat

Google for Startups is promoting a global training program focused on agentic AI. The program teaches founders and developers how to build production-ready autonomous workflows using Google Cloud.

Why it matters

The AI startup center of gravity is shifting from chatbots to systems that can execute tasks across apps, data sources and interfaces. That means founders need more than prompts. They need workflows, grounding, monitoring and enough guardrails to keep the robot intern from emailing the board.

The Deets

• The program covers Gemini Live for real-time voice AI.
• It includes multimodal RAG for grounding AI in data.
• It teaches bidirectional vision agents for data extraction.
• The focus is on moving from demos to production workflows.

Key takeaway

The agent boom is becoming an operations problem, and Google wants founders building that stack on its cloud.

🧩 Jargon Buster - Multimodal RAG: A method that lets AI pull in outside information from formats like text, images or documents before answering.

🧪 Research & Models

Synthetic Brain Scans Enter The Lab

Nvidia released NV-Generate-MR-Brain, a foundation model that generates synthetic 3D brain MRI scans and annotations to support medical imaging AI development.

Why it matters

Medical AI often runs into a data problem: real patient scans are sensitive, expensive and hard to label. Synthetic data can help researchers train and test models while reducing dependence on scarce medical datasets.

The Deets

• The model generates synthetic 3D brain MRI scans.
• It also creates annotations.
• The goal is to accelerate medical imaging AI work.

Key takeaway

Synthetic medical data could help AI researchers move faster without waiting on every real-world scan to clear legal, privacy and labeling hurdles.

🧩 Jargon Buster - Synthetic data: Artificially generated data that mimics real data for training or testing AI systems.

Figure AI’s Robot Books 200 Hours And 249,560 Boxes

Figure AI turned an eight-hour livestream challenge into a 200-hour autonomy marathon, with its Figure 03 humanoid robots sorting nearly 250,000 packages at the company’s Sunnyvale headquarters. CEO Brett Adcock said the run ended with zero hardware failures, a major reliability flex for a company reportedly valued at $39B.

This started as a response to industrial automation veteran Dr. Scott Walter, who challenged humanoid robot makers to prove their machines could survive an eight-hour shift at human speeds without intervention. Figure went 25 times longer, then celebrated with champagne while the robot dubbed ROSE kept sorting boxes like the office party was simply more background noise.

Why it matters

Humanoid robots have had plenty of flashy demos. What they have lacked is proof that they can keep working through the boring, repetitive, physically punishing stuff that real logistics jobs require. Figure’s livestream showed a fleet operating for more than a week, rotating robots to charging docks and continuing the task without a catastrophic mechanical breakdown.

The Deets

• Figure’s unedited autonomous livestream ran for 200 hours.
• The robot fleet processed 249,560 packages.
• Adcock said the run had zero hardware failures.
• The robots used a fleet rotation system, swapping out when batteries ran low.
• Each robot had about four hours of battery life before heading to wireless charging docks built into its feet.
• The autonomy system was powered by Figure’s Helix-02 neural network, which computes actions from raw camera pixels.

Key takeaway

Figure’s demo does not prove humanoids are ready for every job, but it does prove something important: mechanical endurance is getting real. Sorting boxes is narrow work, but a 200-hour autonomous run without hardware failure is the kind of narrow win that can open much wider doors.

🧩 Jargon Buster - Autonomy Stack: The software system that lets a robot perceive its environment, make decisions and move without a human controlling every action.

⚡ Quick Hits

• McKinsey is rethinking billing as AI reduces the value of hourly work and clients push for fees tied to outcomes.
• The White House approved $9B to help U.S. spy agencies acquire advanced AI chips.
• Starbucks scrapped its AI inventory system after nine months because of persistent miscounts and mislabeled products.
• Spotify is launching AI audiobook creation tools with ElevenLabs voices for self-published authors.
• Spotify also struck a Universal Music deal allowing licensed fan-made AI covers and remixes.
• IBM is powering Ferrari’s AI fan app with race summaries, personalization, predictions and an assistant built around F1 data.
• Salesforce is facing scrutiny after Agentforce promotional videos showed AI features and mock-ups that are not widely available.
• YouTube is becoming a larger piracy headache for publishers as AI-narrated audiobook copies spread faster than takedown systems can handle.

🧰 Tools Of The Day

• CData Connect AI: Gives ChatGPT, Claude, Copilot and other AI tools governed read and write access to business data in one unified layer.
• Gemini 3.5 Flash: Google’s new flash model, described by The Rundown AI as 4x faster at half price.
• Polsia: An AI co-founder tool that plans, builds and operates businesses around the clock.
• Perplexity Bumblebee: An open-source scanner for finding risky packages, extensions and AI tool configs during supply-chain incidents.

Today’s Sources: The Internet, The Rundown AI, AI Secret, Humanoids Daily