Forefront

A humanoid robot kept a real factory job. Apollo 2 picked metal brackets on a Mercedes line at 96 percent, per DeepMind. The number is published. How long it kept the job is not.

On November 10, 2026, Google DeepMind announced Gemini Robotics 2.5, its most capable vision-language-action model to date. On active Mercedes-Benz production lines in Kecskemet, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid picked and packed metal automotive brackets with a 96 percent success rate. DeepMind reports the number. It is not independently verified. The job is called kitting: sorting individual parts from bulk bins into organized trays in the exact order each vehicle on the line needs them. The announcement also calls the work a proof of concept, twice.

The question: when does a robot stop being a demo and start being a worker? To answer it, the episode runs one test with three parts: Do, Last, Spread. The rule: a robot has earned the factory floor ONLY if Do, Last, and Spread all pass for that specific robot. Do passes, 96 percent on one task, on a real line. Last is unproven, DeepMind reports no duration, no trial count, no intervention rate, and no failure taxonomy. Spread is claimed but narrow, 200 tasks evaluated in the laboratory, with only one factory job demonstrated. The verdict: on the published evidence, the robot has not yet earned the factory floor.

The verdict: a robot has earned the factory floor ONLY if Do, Last, and Spread all pass for that specific robot. The first passes. The second is unproven. The third is claimed but narrow.

Chapters:

00:00 The question

05:41 Do the job

22:30 Keep doing it

39:17 Do more than one job

56:32 The verdict

If this helped, please like, share, and subscribe, and tell us in the comments: what would convince you a robot is ready for the factory floor, one perfect shift, or a full year without a breakdown?

Watch the full video: https://www.youtube.com/watch?v=EZ8M5MwyxuU

Read the full article: https://forefront1.substack.com/p/forefront-podcast-when-a-robot-keeps

Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xo

Tools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUU

Sponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7Y

Forefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxg

Get in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

Jensen Huang resigned after 33 years at Nvidia. The company he built did $96.2B in a quarter. The question is whether the business stands without its founder.Nvidia reported $96.2B in revenue for the second quarter of fiscal 2027, up 106 percent on the year before. The data-center division alone did $89.0B of it, up 117 percent, about 93 percent of all revenue. Gross margin was 75.0 percent. What Nvidia sells is the whole AI factory: Blackwell accelerators, CUDA software every AI developer learns first, whole data-center systems, and the networking that ties them together. William Blair puts Nvidia at about 85 to 90 percent of AI chips. The order book runs past $1T in combined Blackwell and Vera Rubin purchase orders through 2027, and the guide for this quarter is about $108B, plus or minus two percent. Orders are promises until the cash arrives. The market cut about $600B of value since the investigation began, according to Reuters, while the order book stayed. Colette Kress steps in as interim chief executive, keeping the CFO job she has held since 2013. Analysts at D.A. Davidson report Nvidia has no succession plan in place, and no permanent successor has been named.The question: what does Nvidia actually do, and does the business stand on its own now that the founder who built it is gone? To answer it, the episode runs one test with three parts: Product, Proof, Person. The rule: a founder-built company outlives its founder only if the product, the proof, and the person all stand without him. Product passes, the factory is real and almost no rival can match it. Proof is conditional, the demand is real and the quarter was real, and the guide is still a forecast until the Nov 17 earnings call. Person stays open, 33 years at the top and an empty chair. The verdict: the first stands, the second stands only if the numbers land on the 17th, the third is open.The verdict: a founder-built company outlives its founder only if the product, the proof, and the person all stand without him. The first stands. The second stands only if the numbers land on the 17th. The third is open.Chapters:00:00 The question01:43 Product05:10 Proof08:27 Person10:55 The verdict

Watch on YouTube: https://www.youtube.com/watch?v=jJpEEw5sXFw

Read the full article: https://forefront1.substack.com/p/nvidia-without-jensen-what-it-actually

If this helped, please like, share, and subscribe, and tell us in the comments: what would prove Nvidia is bigger than its founder: the next earnings call, or the next CEO?Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xoTools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUUSponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7YForefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxgGet in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

DeepMind's AI agents found two weaknesses in 1024-bit RSA, with a headline of 1,600x faster. The measured result is 3.5x, and the lock is one nobody should be using anymore.

DeepMind reports, Nov 6, 2026, that two AI agents found two weaknesses in legacy 1024-bit RSA. The headline number is a theoretical 1,600x estimate for signature forgery, arithmetic on paper, never run at 1024 bits. The estimate holds only if the attacker gets to choose a small piece of the data being signed, for example a 32-byte fingerprint. Take away that choice and the estimate no longer applies. The only measured result is 3.5x: the agents reproduced a September 2026 attack on 1024-bit signatures in 391 CPU core-years against 1,380 for the original team, debugging in under 4 days on about 20 short messages from a human. The lock is 1024-bit RSA only. 2048-bit keys and up are not affected, and NIST stopped accepting 1024-bit for new signatures in 2013. The working papers are not peer reviewed, no independent team has replicated the result yet, the experiments ran in an isolated sandbox, and the agents were guided by humans. Two of the people guiding the agents, Heninger and Thomé, wrote the original attack the agents reproduced, so the measured result is a rerun inside one circle.

The verdict: a speedup matters ONLY if Speed, Scope, and Proof all pass, and this one has shown the first only at the measured 3.5x, and neither of the others. The agents did real work, and the work is worth watching. Cheaper reruns change research, even on an old lock. The measured part is real, and it is small. The rest is arithmetic on paper, on a retired lock, with papers still waiting for review. That is why the answer stays mixed.

Chapters:

00:00 The faster it gets, the less it matters

01:08 1,600x faster, nothing broken

03:21 The weaker the lock, the bigger the headline

06:39 The real number is the small one

08:34 Faster is not better

If this helped, please like, share, and subscribe, and tell us in the comments: what would convince you a speedup matters, a bigger number, or a result on a lock people actually use?

Watch on YouTube: https://www.youtube.com/watch?v=etULkTyqJmM

Read the full article: https://forefront1.substack.com/p/deepmind-rsa-1600x-faster-on-a-lock

Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xo

Tools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUU

Sponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7Y

Forefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxg

Get in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

A robot called Theo restocks supermarket shelves at 400 items an hour, overnight, while the store is closed. One store in Zurich, one launch film, and no independent proof yet.

Theseus Robotics reports 400 units an hour in its own launch film from November 7, 2026, shown as on-screen graphics. That is the company's claim about its own robot, and no independent test has confirmed it. The store is a SPAR in Zurich. It closes at 9pm, and Theo works the night, with no staff in the building. By day, Theo picks online orders, up to 40 items an hour. Two stores so far, one in Zurich, and only the Zurich store is in the film. Retailer pilots planned for 2027. No price. No error rate. Theseus uses remote human control initially to train the robots.

The verdict: rate claimed but not verified, range partly shown, proof missing. A robot has earned the night shift only if Rate, Range, and Proof all pass for that specific robot.

Chapters (verified against audio manifest v2, 50 turns, 722.93s):
00:00 The question
01:05 Rate
03:54 Range
06:56 Proof
09:38 The verdict

If this helped, please like, share, and subscribe, and tell us in the comments: would you trust a robot to restock your local supermarket overnight, with no staff in the building?

Watch on YouTube: https://www.youtube.com/watch?v=sILuH-VmqQA

Read the full article: https://forefront1.substack.com/p/theo-the-supermarket-robot-400-an

Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xo
Tools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUU
Sponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7Y
Forefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxg
Get in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

Amazon is hiring for a new home robot, five years after Astro stalled at a few thousand units. The money moved first: $33.4B into robotics in six months, more than all of 2025.PitchBook reports $33.4B went into robotics and physical AI in the first half of 2026, more than the $28.7B of all of 2025. Astro cost $1,450, sold invite-only, and stalled at a few thousand customers. The business version died within a year. There is no product yet, no name, no price, no date. The receipt is the hiring and the funding, not a deployment. Fauna's Sprout is a separate humanoid, $50,000 Creator Edition, not the new robot.The verdict: money passes, home unproven, proof missing. A home robot earns its place only if Home, Money, and Proof all pass for that specific robot.Chapters:
00:00 The question
01:21 Home
03:56 Money
06:40 Proof
09:22 The verdict
Watch on YouTube: https://www.youtube.com/watch?v=RaYe-3Rm4us
Read the full article: https://forefront1.substack.com/p/amazons-new-home-robot-the-money
If this helped, please like, share, and subscribe, and tell us in the comments: what would a home robot have to do to earn a place in your home?Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xoTools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUUSponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7YForefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxgGet in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

how Claude dreams: between sessions, Claude replays its logs, corrects wrong notes, and builds a wiki. The agent that wrote the most remembered the least.

Anthropic reports that in Pokemon Red, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned 1 badge, at about 10,000 steps into the game. Claude Fable 5 wrote nine files in two folders, plus a single top-level file called global.md, and earned 4 badges. Same game, same rough point, one run each, no independent verification yet. Then came dreaming: an offline pass between sessions in which Claude reflects on its logs to fix wrong notes and build linked representations such as wikis. It does not retrain the model or change weights. Anthropic reports that on a wiki test over a corpus of about 350,000 files, roughly 420M tokens, wiki accuracy grew from 13% to 48% of graded criteria on 25 questions as dreaming scaled from 2 to 128 sessions.

The verdict: judge an agent's memory by whether its notes stay few, tested, and corrected against the record, not by how many notes it writes. Memory is a prediction about what the next session will need, not a recording of everything that happened.

Chapters:

00:00 The question

01:05 Remember everything

03:00 Remember almost nothing

05:35 Dreaming

10:22 The verdict

If this helped, please like, share, and subscribe, and tell us in the comments: what should an agent remember between sessions, and what should it forget?

Watch on YouTube: https://www.youtube.com/watch?v=94YMDt6rvgw

Read the full article: https://forefront1.substack.com/p/how-claude-dreams-memory-between

Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xo

Tools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUU

Sponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7Y

Forefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxg

Get in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

when AI starts improving AI: oversight holds ONLY if Notice, Check, and Steer all pass for that specific capability. AI is now improving AI, and the question is no longer how smart it gets, it is whether anyone can still check its work.

This week, three events from Nov 2 to 4, 2026 put that question to the test. Anthropic told the SEC it may imminently meet its own definition of dramatically accelerated progress attributable to AI, and admitted it may not spot a significant acceleration until after it has occurred. MIT showed GPT-6 Astra editing another model's weights directly: 131,072 weights, 0.016 percent of Qwen3.5-0.8B, moving MATH-500 from 7.9 percent to 48.1 percent in the authors' run. And OpenAI's pipeline produced 948 working Linux root exploits for $1,213,430, about $1,279 per distinct exploit.

The verdict: oversight holds ONLY if Notice, Check, and Steer all pass for that specific capability. On the evidence from Nov 2 to 4, it has not been demonstrated to keep pace.

Chapters:
00:00 The question
01:02 Notice
22:55 Check
32:55 Steer
46:43 The verdict
50:31 What to watch

If this helped, please like, share, and subscribe, and tell us in the comments what you would trust an AI to change without a human checking it first.

Work with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xo
Tools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUU
Sponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7Y
Forefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxg
Get in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewform

Read the full article: https://forefrontshowofficial.substack.com/p/forefront-podcast-when-ai-starts

Watch the full video here: https://www.youtube.com/watch?v=9uxvl2Hwg-U

The most expensive lesson in robotics is the one you have to teach every robot, one by one. Can a robot do a task it was never shown on a robot? This episode runs three tests on one claim: Watch, Transfer, and Attempt. Watch asks whether a robot can learn a job from human experience instead of robot demonstrations of that job. Transfer asks whether the skill survives the move to a different body. Attempt asks whether it actually does jobs nobody demonstrated, and how often. The verdict is one rule: a robot has learned a job it was never shown ONLY if Watch, Transfer, and Attempt all pass for that specific job.The claim comes from LiberAI, announced November 3, 2026. Liber-0 is a World Action Model pretrained on more than 2.2 million hours of human experience, and LiberAI reports it learned from human video, glove-recorded motion, and robot data. LiberAI reports 36.0 percent success across 60 tasks nobody demonstrated on a robot, 600 trials in total. Every Liber-0 number in this episode is vendor-reported, and no independent verification exists yet. The Robot Report piece is marked sponsored content by LiberAI, and the press release carries a disclaimer that it has not been independently verified.Chapters:00:00 The question01:08 Watch03:17 Transfer06:29 Attempt09:57 The verdictListen to the full episode here on Spotify.Follow Forefront on X: https://x.com/Forefrontshow_Follow Forefront on Instagram: https://www.instagram.com/forefrontshow/Follow Forefront on LinkedIn: https://www.linkedin.com/company/forefrontshowRSS feed: https://anchor.fm/s/11810b5d0/podcast/rssWork with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xoTools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUUSponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7YForefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxgGet in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewformKeywords: how a robot learns a job it was never shown, Liber-0 explained, zero-shot robot manipulation, LiberAI, LiberSense Glove, robot learning human video, world action model, physical AI.

Watch on YouTube: https://www.youtube.com/watch?v=bnJm4Lj9hWY

Read the full article: https://forefrontshowofficial.substack.com/p/how-a-robot-learns-a-job-it-was-never

A model that remembers everything would be useless, and a model that remembers nothing is already here. What should an AI system remember, forget, and prove it remembered correctly? This episode runs three tests on one long project, over weeks: Store, Recall, and Forget. Store asks whether a fact can be written down with a source and a time so it survives the session. Recall asks whether the right memory comes back, with its source attached, and is used to act correctly. Forget asks whether a fact can be let go on purpose, with an audit trail, and proved to stay gone. The verdict is one rule: useful memory only if Store, Recall, and Forget all pass for that specific task. Evidence includes retrieval-augmented generation (Lewis and colleagues, 2020), MemGPT (Packer and colleagues, 2023), Generative Agents (Park and colleagues, 2023), and the Lost in the Middle position effect (Liu and colleagues, 2023, 75.8 percent at the start of a 20-document set, 53.8 percent in the middle, 56.1 percent with no documents; a 2023 study, not a current product score).Listen to the full episode here on Spotify.Follow Forefront on X: https://x.com/Forefrontshow_Follow Forefront on Instagram: https://www.instagram.com/forefrontshow/Follow Forefront on LinkedIn: https://www.linkedin.com/company/forefrontshowRSS feed: https://anchor.fm/s/11810b5d0/podcast/rssWork with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xoTools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUUSponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7YForefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxgGet in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewformKeywords: how AI memory works, AI agent memory, LLM memory, RAG memory, context window vs memory, AI forgettingWatch on YouTube: https://www.youtube.com/watch?v=m_-u2_6OER8Read the full article on Substack: https://forefrontshowofficial.substack.com/p/how-ai-memory-works-store-recall

A robot that is safe in a cage is not safe beside you. In a shop aisle, a worker restocks a shelf while a scanning robot rolls past, and a shopper steps in with a child beside her. Can that robot be trusted to share a workspace with a person? This episode tests two applications only, retail shelf-scanning in a shop aisle and a cobot arm in a shared cell, against three tests: Prove, Contain, and Repeat. Prove asks for a bench test a third party can run again. Contain asks whether a separate safety monitor holds when the main system fails. Repeat asks whether safety is tested again after change, with a log kept and a named human able to stop the system.Listen to the full episode here on Spotify.Follow Forefront on X: https://x.com/Forefrontshow_Follow Forefront on Instagram: https://www.instagram.com/forefrontshow/Follow Forefront on LinkedIn: https://www.linkedin.com/company/forefrontshowRSS feed: https://anchor.fm/s/11810b5d0/podcast/rssWork with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxpxxwxy7xoTools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUUSponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elk/GFZdywG2nC0QOsHjz8AaLySz7N1gNZ-u26OUCX9Ks7YForefront job board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxgGet in touch: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewformKeywords: robot safety, collaborative robots, ISO 10218, ISO/TS 15066, safety-rated monitored stop, speed and separation monitoring, power and force limiting, UL 3300, UL 1740, PLd Category 3, bench test, safety monitor, re-test log, Prove, Contain, RepeatWatch the full video on YouTube: https://www.youtube.com/watch?v=V-Wanzye-2oRead the full article on Substack: https://forefrontshowofficial.substack.com/p/how-a-robot-proves-it-is-safe-to