The week in AI: Oct. 26-⁠31, 2026

Step-by-step 3D assembly guides for IKEA furniture

On Oct. 27, Hassan El Mghari, director of developer experience at Together AI, and designer Youssef launched Bygga, a free, open-source website under the MIT license. Users type the product name or article number from an IKEA box, and the site shows the assembly in step-by-step 3D, drawn in the style of the paper manual, down to every screw and tool. The site lists 1,223 products; its GitHub repository counts 1,171 products in 2,180 sizes from IKEA’s UK and US ranges.

According to the repository, the guides come from a parametric plan generator built from IKEA’s public listings, not from a generative model. Steps are read aloud in English from 1,476 prerecorded clips, and a Whisper-based listener moves to the next step on the word “next.” Bygga is not affiliated with IKEA, and every guide links to IKEA’s official manual.

Hassan@nutlope

Introducing Bygga!

It walks you through assembling IKEA furniture step by step, with interactive 3D & voice instructions.

100% free & open source.

Simulon adds digital characters to scenes shot on a phone

On Oct. 28, Simulon founder Divesh Naidoo announced Reality-Aware Rendering in the company’s iPhone app. Users place 3D “neural assets,” generated from photos, sketches or text, in a real location and then act out the scene while recording. According to Naidoo, when recording stops, Simulon reads everything in the take, including what the person does and says and the place around them, and the assets respond to all of it.

The result is not a live overlay. Naidoo says the final take is regenerated from the recording, a new performance of the scene rather than a copy of the footage. Simulon’s website says processing happens in the cloud. The free tier includes 100 starter credits, while creating custom assets requires the $14.99-a-month Plus plan or the $39.99 Pro plan. An Android version is “coming soon.”

Divesh Naidoo@diveshnaidoo

Introducing Reality-Aware Rendering in @Simulon, a completely new way to create with the real world.

Place neural assets with your phone and act out the scene. When you stop recording, Simulon renders the take, reading everything in it (what you do, what you say and the place around you), and your assets respond to all of it.

The final is regenerated from your recording: a new performance of the scene, not a copy of the footage.

Available now on iPhone. Coming to Android soon.

Glasses that turn silently mouthed words into text

On Oct. 30, The Subvocal Company, part of Y Combinator’s Fall 2026 batch, said it had raised $1.3 million from YC and Afore Capital. According to founder and CEO Aarush Yadav, its glasses track internal speech muscles with non-contact sensing and decode silent, almost invisible mouthing into text at more than 200 words per minute. The company says the glasses send radar signals whose reflections shift with tiny speech-muscle movements, and an AI model turns those shifts into text.

The 200-word figure is the company’s own prototype claim, with no independent accuracy data. A Georgia Tech article from September said the startup’s earlier letter-level prototype spelled out words at 30 words per minute. Pre-orders take a refundable $199 deposit toward a $499 Early Bird price. US-only shipping is estimated for the first quarter of 2027, subject to FCC approval.

Aarush Yadav@AarushYadav47

We’ve raised $1.3M from YC & Afore Capital so you never have to type or speak again.

Typing is slow. Talking into your computer is awkward. Subvocal fixes both

Our glasses give you the speed of speech without you making a sound or touching a keyboard.

It tracks your internal speech muscles with non-contact sensing and decodes silent, almost invisible mouthing into text at more than 200 words/min.

Pre-orders are now live at:
subvocal.co/preorder?utm_source=x

NVIDIA opens up MotionBricks, a real-time motion model

MotionBricks, a SIGGRAPH 2026 paper by NVIDIA researchers, generates motion in real time for game characters and robots. A single model learns from more than 350,000 motion capture clips of real actors. Its “smart primitives” interface combines gait styles such as zombie or limping with interactions such as picking up a sword or sitting on a bench, without retraining. On Oct. 23, NVIDIA opened a standalone repository with the code plus a human (SOMA) model and a Unitree G1 robot model; it went public around Oct. 26.

According to the paper, the model produces 15,000 motion frames per second on a desktop with an RTX 5090 and fills in motion between two key poses with 2 ms of latency. Those frames are 30 Hz skeleton pose data, not rendered video. The code is Apache 2.0; the weights use the NVIDIA Open Model Agreement, which allows commercial use under conditions.

TBC releases a video model it says was derived from neurons

On Oct. 30, San Francisco’s The Biological Computing Co. (TBC) released mischief 1, an image-to-video model, on fal and OpenRouter. It is built on Wan 2.2, Alibaba’s open-weight video model. TBC grows neurons in the lab, measures how they process information and turns the computational principles it finds into software. The company says the neurons stay in the lab, and the added layer increases the model’s parameter count by less than 0.1% and runs on standard hardware.

TBC claims mischief 1 matches Wan 2.2’s quality while generating video 5x faster; no independent test has been published. On fal, 480p video costs half a cent per second, or 2.5 cents for a 5-second clip, and 720p costs twice that. In mid-October, TBC’s Neural Dynamics Adapter, a software prototype based on its neuron research, made TIME’s Best Inventions of 2026 list.

Alex Ksendzovsky@AlexKsendzovsky

today, we introduce a little mischief to the world.

for four years, we’ve been brewing real brain cells on electrodes to understand how they compute. now, we’ve turned those discoveries into a video model anyone can use.

from 0.5¢ a second, mischief 1 is at least 5x faster than its base and the cheapest image-to-video model on @fal and @OpenRouter. soon to be on @awscloud marketplace.

it’s not a delusion anymore. biocompute is real, commercial and this is just the beginning.

happy halloween.

Google DeepMind unveils Genie 4, a world model that also generates sound

Announced by Google DeepMind on Oct. 27, Genie 4 builds a real-time, interactive world from a text prompt, which users explore with a keyboard. DeepMind calls it its first real-time world model to generate audio natively, producing ambient sound, music and characters’ speech along with the visuals. Users can change the weather or add a new person or object with a text prompt. Gemini 3.8 Flash acts as a “guiding agent,” watching the scene and breaking goals into steps.

According to TechCrunch, a scene can stay consistent for up to two minutes, double its predecessor’s limit. For now, Genie 4 is open only to trusted testers. In the coming weeks, it will reach Google AI Ultra subscribers in Project Genie and early access partners on the Gemini API. DeepMind acknowledges that visual details and object permanence can drift in longer sessions.

Google DeepMind@GoogleDeepMind

This is Genie 4, our most realistic and interactive world model yet. 🌐

Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.

Altman: scarce memory is better used in the cloud than in PCs

On Cheeky Pint, the podcast hosted by Stripe co-founder John Collison and released Oct. 29, OpenAI CEO Sam Altman addressed the push to run AI locally. Many people, he said, assume local open-source models will remove the need for data centers. “From a global allocation perspective, actually, it’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer,” he said, used about 5% of the time. In such a supply-constrained environment, he argued, memory should run at maximum utilization in the cloud.

Tom’s Hardware noted that OpenAI’s main business is closed models running in data centers, so Altman was effectively defending the company’s business model. In October 2025, OpenAI signed letters of intent with Samsung and SK hynix to supply memory for Stargate. That announcement cited up to 900,000 DRAM wafers a month.

The first reply from Google’s TPU in orbit was a haiku

On Oct. 30, Google said it had received the first AI response from its Project Suncatcher prototype satellite, built with Planet and launched on Oct. 1. The team asked the onboard Gemma 4 open model for a haiku about being in space. The reply was downlinked as the satellite passed over Svalbard, Norway. Google calls it the first AI response from a data center class TPU in orbit. It is not the first satellite to run an AI model: Starcloud-1 ran Gemma in orbit in 2025.

According to TechCrunch, only one of the satellite’s four TPUs can run at a time. The satellite generates up to 1 kW, and a TPU can stay on for about 20 minutes at a stretch. Thermal and radiation tests come next. Google plans to launch two more satellites, linked by optical connections, in 2027.

NVIDIA releases Nemotron 3.5 Transcribe, an open-weight speech model

On Oct. 28, NVIDIA released Nemotron 3.5 Transcribe as an open-weight model on Hugging Face. The 30-billion-parameter mixture-of-experts (MoE) model, with about 4 billion parameters active per token, combines transcription, speaker attribution, translation and audio question answering in one checkpoint. It transcribes 40 languages and translates 31 of them to or from English. In 13 languages, it labels who said what for up to eight speakers, even when people talk over each other. It handles recordings up to an hour long and is built for recorded audio, not live streaming.

NVIDIA says it ranks first among open-weight models on the English track of the Hugging Face Open ASR Leaderboard, with a 5.33% average word error rate. Turkish transcription is rated “production-ready,” but Turkish translation and multispeaker transcription are not supported. The OpenMDW-1.1 license allows commercial use.

NVIDIA AI@NVIDIAAI

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️

An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.

Watch it keep track of who said what when everyone’s talking at once.

1/ 🧵

Anthropic acquires Mithril Medicine, a one-person drug company

On Oct. 27, Anthropic said it had acquired Mithril Medicine, a one-person company; the accompanying technical report says the deal closed in July. Founder Cody Tranbarger used Claude to organize contracts, logistics and experiments with 11 outside labs. Mithril is developing an antibody that bridges two proteins for LAMA2-related congenital muscular dystrophy, a rare childhood muscle disease. In a six-week pilot study, mice given the tetravalent antibody developed about twice the peak force and grip strength of vehicle-treated mice.

The report is not peer reviewed, and the results are in mice only. It says that in one week Claude drafted 17 emails, the team sent 12 after editing, and at least three errors reached contract research organizations. Anthropic expects human trials to start as soon as late 2028. On X, Tranbarger described this way of working as “Pipeline-in-a-Person.”

Professor Oak@Prof_Oak_

Surprise 1: Anthropic acquired a one-person biotech
.
.
.
Surprise 2: that person is me
Surprise 3: I’m here to tell you all about it

Yes, this is what you think it is; after 6 yrs of increasingly loose anonymity: the Professor Oak reveal post

…and yes, it has gotten so boring around here that I’m literally doxxing myself to spice up the timeline

-------
Ok, allow me to introduce myself. My name is Cody Tranbarger. I mostly tweet, but occasionally find time to dabble in drug dev & starting/building biotechs. My past lives span the VC, crossover, startup continuum; Prof Oak has been my way of keeping a digital foot in those worlds, despite having my physical feet planted elsewhere.

The past 18 mo, ‘elsewhere’ was Mithril Medicine, a co. I founded to develop bsAbs for LAMA2-CMD & other ECM-related genetic diseases.

Mithril was born out of hands-on building with @MikeNGladstone, @JasonPaulRhodes, @LifeSciVC, & team @atlasventure. Somehow, I made enough half-decent calls along the way (even a broken clock is right twice a day, eh?) to convince them Mithril/I warranted a small seed. While I can hardly rationalize such an outrageously cavalier bet, I’m grateful for it, & I'd partner with Atlas again in a heartbeat.

In typical AV fashion, Mithril’s primordial form was virtual, decentralized, & ultra-lean. This stage tends to be short-lived, but for Mithril, it worked surprisingly well, for surprisingly long, thanks in part to the rapid maturation of the AI tools I used liberally.

Eventually, Mithril became a bit of an experiment itself - to find the upper bound of how far it could scale as a single operator + a bunch of Claude (+ a stellar bench of part-time consultants).

In parallel, I started spending time with @nc_frey, @an1lam, & @CoefficientBio team; and promptly experienced, for the 1st time, AI at the frontier. Soon thereafter, Mithril stopped looking like an experiment, & started looking like a preview of an entirely new operating model.

From that spark came “Pipeline-in-a-Person.” The idea immediately resonated when we shared it here back in April; and it was central to the shared vision that led to Anthropic’s acquisition of Mithril, & my joining the team.

Why join Anthropic? Three reasons.
(1) For Mithril, the programs continue, with more resources & infrastructure behind them, and as much Claude as I can put to work.

(2) For me, the chance to start several more programs, & find out how many one person can actually run in parallel - an almost unprecedented luxury for a curious drug hunter.

(3) For our mission, a path to institutionalize & scale this new way of working, then put it in the hands of all of biopharma, all over the world - which, candidly, I find the most energizing.

Will AI x Bio cure diseases? Unequivocally yes, imo… but not by the route that gets the most airtime. Rather than timeline acceleration & compression, I see orchestration & parallelization as today's low-hanging fruit.

Most drug hunters see a handful of go/no-go calls in a career, & spend entirely too much of the rest trading emails (my inbox can attest).

Give them the bandwidth to run several programs at once, and the time to focus on scientific judgment, and I’m willing to bet they’ll make better calls. And at scale, better calls translate to more drugs, higher hit rates, broader patient impact, & likely, a disease or 10 or 100 cured. I couldn't be more excited to build toward that future @AnthropicAI.

Finally, a heartfelt thank you to BioTwitter/X. The end of anonymity is bittersweet, but if anyone’s earned a proper introduction, you all have.

It’s difficult to overstate how much this app has done for me: a decade of education, plenty of friends, many job opportunities, several mentors, a few investors, even a co-founder (fmr). It was the generosity of this community that inspired me to start posting 6 years ago, after just as many years lurking. I hope I’ve been able to pay it forward, in my own small way.

But, to be clear, Prof Oak isn’t going anywhere. Same beat, same themes, and now, a front-row seat to drug development at the frontier. Stay tuned for dispatches from the trenches; and remember...

Gotta Cure 'Em All

A gallery of ready-made legal tools for Claude for Legal

On Oct. 28, Anthropic added a gallery of ready-made tools, which it calls “artifacts,” to Claude for Legal. Anthropic’s six tools include a cite check, a playbook deviation board that compares contracts with a company’s standard terms and a regulatory horizon scan of upcoming rules. A tabular review that pulls chosen terms from a set of contracts into a table is another. Goodwin, Harvey, Foley & Lardner and PwC added tools of their own, and the gallery also has nine access-to-justice tools.

Each tool opens with a built-in example, and lawyers can add their own documents and connectors and share it across the firm. In the workflow Anthropic describes, Claude does the first pass and a lawyer reviews the results. The announcement does not say which plans include the gallery, and no independent accuracy measurement has been published.

Claude@claudeai

New in Claude for Legal: a gallery of ready-made artifacts for contract review, regulatory horizon scans, cite checks, and more.

Start from the built-in example, add your own documents and connectors, and share it with anyone across the business.

Reddit stock falls 29% as search traffic slips

Reddit reported third-quarter results on Oct. 29: revenue rose 53% to $895 million, and net income reached $283 million. In his shareholder letter, though, CEO Steve Huffman wrote that search referral traffic fell meaningfully in the quarter, with the decline accelerating toward its end. According to the letter, Reddit lost more than 1 million referrals per week on average in October. The letter does not name a search engine. On the earnings call, according to The Verge, Huffman said Google’s AI Mode “does not drive traffic.”

Huffman said daily users could fall by a “low-double-digit” percentage in the fourth quarter if the trend continues. Reddit, largely readable without an account for about two decades, will require logins, starting with tests in select markets outside the US in the fourth quarter. According to The Verge, the stock fell 29% on Oct. 30.

China’s STAR50 index loses 47% in four months

The STAR50, made up of the 50 largest stocks on Shanghai’s STAR Market tech board and dominated by chip and AI companies, closed Oct. 30 at 1,160.84. That is down 47.4% from its June 30 peak of 2,207.86 and its lowest close since Aug. 21, 2025, wiping out all of its 2026 gains. A Reuters analysis on Oct. 26 said the selloff had erased about $480 billion in market value by then.

According to Reuters, there was no obvious selling trigger. Traders cited high prices, tax bills and the threat of the US or other nations banning Chinese equipment in data centers. The index’s price-to-earnings ratio had neared 250 at its peak. No new US ban has taken effect. BNP Paribas strategist William Bratton still sees “buying opportunities for the careful.”

Karpathy builds his own YouTube from videos recommended on Hacker News

On Oct. 31, Andrej Karpathy introduced HackerTube on his blog. Opus 5.5 in Claude Code downloaded every Hacker News comment and had Sonnet label them. After spam was filtered out, about 40,000 comments recommending a YouTube video yielded 6,454 videos, recommended between 2007 and 2026 and spread across 150 topics. The site presents them in a YouTube-style interface with a recommendation feed that needs no account.

Karpathy says the project took about 2 hours; he only got an API key, started the server and bought the domain. In a Hacker News comment, he said the work used 504 million input tokens and 7.6 million output tokens, more than 90% of them on Sonnet calls. Karpathy has worked at Anthropic since May, so the assessment comes from an employee of the model’s developer.