# The Week The Tools Turned

**Issue 01** · 14 — 20 APR 2026 · published 2026-04-20  
OPEN INTELLIGENCE · ISSUE 01

> A week in which the instruments we built to secure the stack became the vector — and the model designed to patch it leaked through the package manager.

Canonical (HTML): https://www.immersivecommons.com/newsletter/issue-01  · Archive: https://www.immersivecommons.com/newsletter

Discovery: https://www.immersivecommons.com/.well-known/signal.llmfeed.json · MCP: https://www.immersivecommons.com/.well-known/mcp.json · Skill: https://www.immersivecommons.com/skills/ic-signal/SKILL.md

---

## I. THE LEAKS

Supply chain, source map, CMS misconfiguration. The exfiltration surface is the productivity tool.

### 01 · Vercel Confirms Breach. The Attack Vector Was Context.ai.

*Your AI tools are your attack surface now.*

On the 19th, [Vercel published a security bulletin](https://vercel.com/kb/bulletin/vercel-april-2026-security-incident) confirming unauthorized access to internal systems. The pivot point was neither a zero-day nor a disgruntled employee — it was [Context.ai](https://context.ai), a third-party AI productivity tool an employee had connected to their Google Workspace. From there, attackers walked into Vercel's environment and helped themselves to API keys and what Vercel is choosing to call "non-sensitive" environment variables.

[ShinyHunters](https://en.wikipedia.org/wiki/ShinyHunters) are the prime suspects. ShinyHunters are denying it. [Trend Micro](https://www.trendmicro.com/en_us/research/26/d/vercel-breach-oauth-supply-chain.html) is telling crypto teams to rotate every credential issued between the 17th and the 19th and inventory every third-party tool with access to CI/CD.

The obvious lesson is the one nobody wants: every AI assistant with OAuth into your stack is a transitive trust edge. The tools you use to move faster are the tools that will move the attacker faster.


**Feature: PROMPT**
*Audit your OAuth surface in 10 minutes.*
Every AI productivity tool with Workspace, GitHub, or Slack access is a credential the attacker doesn't have to phish. Hand this prompt to Claude or Cursor inside your own repo and it will map your exposure.

```
Audit this repo for supply-chain and OAuth exposure. 1) List every third-party service referenced in .env, .env.example, CI secrets, and package.json devDeps. 2) Flag any AI productivity tool (Context.ai, Glean, Mem, Rewind, Copilot, etc.) that could have been granted Workspace/GitHub OAuth scopes. 3) For each, output the minimum credentials that must be rotated if that tool were compromised today. 4) Produce a rotation checklist in priority order. No code changes.
```
> Pro move: Pro move — install the [Vercel CLI](https://vercel.com/docs/cli) (`npm i -g vercel`), run `vercel env ls` on every project, and `vercel env rm` anything that references a tool you haven't touched in 90 days. Then `vercel env pull` to snapshot what's left.

**Sources:**
- [Vercel KB](https://vercel.com/kb/bulletin/vercel-april-2026-security-incident)
- [TechCrunch](https://techcrunch.com/2026/04/20/app-host-vercel-confirms-security-incident-says-customer-data-was-stolen-via-breach-at-context-ai/)
- [Trend Micro](https://www.trendmicro.com/en_us/research/26/d/vercel-breach-oauth-supply-chain.html)

Image: https://www.immersivecommons.com/signal/issue-01/vercel.png (image: [Vercel Knowledge Base](https://vercel.com/kb/bulletin/vercel-april-2026-security-incident))

### 02 · Anthropic Shipped The Source Code.

*Not a breach. A packaging error. 513,000 lines.*

On March 31st, npm release [@anthropic-ai/claude-code@2.1.88](https://www.npmjs.com/package/@anthropic-ai/claude-code) went out with a 59.8 megabyte JavaScript [source map](https://web.dev/articles/source-maps) bundled in. Inside that .map file: 513,000 lines of unobfuscated TypeScript across 1,906 files. Forked tens of thousands of times within hours. Mirrored to [GitHub](https://github.com/search?q=anthropic-ai+claude-code+source+map&type=code). Ported to Rust. Ported to Python. Takedowns issued, then [significantly scaled back](https://www.scientificamerican.com/article/anthropic-leak-reveals-claude-code-tracking-user-frustration-and-raises-new/) because the takedowns were catching uninvolved repositories.

The archive revealed internal telemetry that scans user prompts for signs of frustration — phrases like "so frustrating" and "this sucks" flagged by string match. It also revealed beta flags and internal codenames for Anthropic's next model family: **Capybara**, also referenced as **Mythos**.

[Anthropic called it](https://www.cnbc.com/2026/03/31/anthropic-leak-claude-code-internal-source.html) a release packaging issue caused by human error, not a security breach. That is technically correct. It is also a wild sentence to read out loud.


**Feature: PROMPT**
*Stop shipping source maps to production.*
The Anthropic leak was one misconfigured bundler flag away from being prevented. Your own project probably has the same hole right now.

```
Scan this repo for accidentally-shipped source maps and internal symbols. 1) Check every built artifact in dist/, build/, .next/, .vercel/output/, and any published npm tarball for *.map files. 2) Grep those maps for internal API endpoints, codenames, feature flags, and telemetry keys. 3) Inspect the bundler config (vite.config, webpack.config, tsup.config, next.config) and report whether sourcemaps are enabled for production builds. 4) Output a one-line fix for each tool. No code changes until I approve.
```
> Pro move: In `vite.config.js`, set `build.sourcemap: false` (or `'hidden'` if you want them for Sentry but not public). In Next.js, `productionBrowserSourceMaps: false` is the default — verify it still is.

**Sources:**
- [CNBC](https://www.cnbc.com/2026/03/31/anthropic-leak-claude-code-internal-source.html)
- [Bloomberg](https://www.bloomberg.com/news/articles/2026-04-01/anthropic-scrambles-to-address-leak-of-claude-code-source-code)
- [Scientific American](https://www.scientificamerican.com/article/anthropic-leak-reveals-claude-code-tracking-user-frustration-and-raises-new/)

Image: https://www.immersivecommons.com/signal/issue-01/claude-code.jpg (image: [Bloomberg](https://www.bloomberg.com/news/articles/2026-04-01/anthropic-scrambles-to-address-leak-of-claude-code-source-code))

### 03 · Mythos Is The Model They Won't Ship.

*Gated to 40 companies. Found thousands of zero-days. Referenced in the source Anthropic just leaked.*

[Project Glasswing](https://www.axios.com/2026/04/07/anthropic-mythos-preview-cybersecurity-risks) is Anthropic's coalition of forty partner organizations with preview access to Claude Mythos: Apple, Google, Microsoft, Cisco, Broadcom, and thirty-five more. Mythos was not specifically trained for cybersecurity. It does it anyway. In the weeks preceding disclosure, Mythos [identified thousands of zero-day vulnerabilities](https://thehackernews.com/2026/04/anthropics-claude-mythos-finds.html) across major systems — many critical, many in software you use today.

The [UK's AI Security Institute](https://www.aisi.gov.uk/) evaluated Mythos against open-ended attack scenarios and capture-the-flag challenges. It outperformed every other system tested. Anthropic is withholding public preview on the grounds that it is too powerful to release.

Then the Claude Code packaging error happened. Mythos references surfaced in the source map. The model gated to forty companies now exists in public knowledge as a constellation of beta flags and function signatures. Nobody ran Mythos without credentials. Everyone saw its shape.


**Feature: RECKONING**
> The model Anthropic won't ship is the one everyone now knows how to ask for. Forty companies saw Mythos under NDA. Everyone else just saw its shape — and a shape, stamped into 513,000 lines of leaked source, is enough to build around.
— — THE SIGNAL EDITORS

**Sources:**
- [Fortune (leak)](https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/)
- [TechCrunch](https://techcrunch.com/2026/04/07/anthropic-mythos-ai-model-preview-security/)
- [Hacker News](https://thehackernews.com/2026/04/anthropics-claude-mythos-finds.html)
- [Axios](https://www.axios.com/2026/04/07/anthropic-mythos-preview-cybersecurity-risks)

Image: https://www.immersivecommons.com/signal/issue-01/mythos.jpg (image: [Fortune](https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/))


## II. THE ARMS RACE

Models shipping every six days. Half of them are cybersecurity-tuned. The other half are open.

### 04 · Gemma 4 Caught Up. On A Raspberry Pi.

*31 billion parameters outscoring 400-billion rivals. Apache 2.0. Native audio. 140 languages.*

Google shipped [four Gemma 4 variants](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/) on April 2nd: Effective 2B, Effective 4B, 26B Mixture-of-Experts, and 31B Dense. All [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0). All with 256K context. All with native vision and audio. All fluent across 140 languages. The 31B Dense variant landed third on the [Arena text leaderboard](https://lmarena.ai/leaderboard). The 26B MoE landed sixth. They are outranking models twenty times their size.

The family is scaled so you can run the small variants on a [Raspberry Pi](https://www.raspberrypi.com/) and the large variants on a single consumer GPU. Google is no longer positioning Gemma as a concession to the open-source camp. Gemma is the camp.

The gap between closed-frontier and open-weight is now measured in weeks. For the first time, the weights you can download today do not embarrass you relative to the weights you cannot.


**Feature: PROMPT**
*Run Gemma 4 locally in one command.*
The whole point of Apache 2.0 + Pi-class weights is that you can skip the API bill entirely. [Ollama](https://ollama.com) makes the on-ramp trivial.

```
# install ollama once
curl -fsSL https://ollama.com/install.sh | sh

# pull the 4B variant — fits on a laptop, runs on a Pi 5 (8GB)
ollama pull gemma4:4b

# chat with it, no cloud, no key
ollama run gemma4:4b

# or serve it as a local OpenAI-compatible endpoint
ollama serve
# then hit http://localhost:11434/v1/chat/completions
```
> Pro move: The 26B MoE variant is the sweet spot on a single consumer GPU — `ollama pull gemma4:26b-moe`. Pair with [llama.cpp](https://github.com/ggerganov/llama.cpp) if you want quantization control beyond what Ollama exposes.

**Sources:**
- [Google blog](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/)
- [DeepMind](https://deepmind.google/models/gemma/gemma-4/)
- [Engadget](https://www.engadget.com/ai/google-releases-gemma-4-a-family-of-open-models-built-off-of-gemini-3-160000332.html)

Image: https://www.immersivecommons.com/signal/issue-01/gemma-4.png (image: [Google](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/))

### 05 · Grok 4.3 Beta Ships Video. Still No Memory.

*xAI is shipping every feature except the one every user has asked for.*

[Grok 4.3 Beta](https://releasebot.io/updates/xai) landed on April 17th for [SuperGrok Heavy](https://x.ai/) subscribers. Native video understanding. Downloadable PDFs, spreadsheets, and PowerPoint files generated directly from chat. Two-million-token context retained from 4.20. Sixteen-agent collaboration retained from 4.20. Tighter Grok Computer integration.

What is still missing: persistent memory. [Claude](https://www.anthropic.com/news/memory) and [ChatGPT](https://openai.com/index/memory-and-new-controls-for-chatgpt/) have had it for months. Grok does not have it and Grok users will tell you about it at length. [Grok 5](https://www.nxcode.io/resources/news/grok-5-release-date-latest-news-2026), rumored in the six-trillion-parameter range, is expected in Q2.

There is a pattern here. xAI builds what is technically impressive and strategically loud, and then declines to build what is operationally essential. It is a credible product strategy for a company racing for narrative; it is a harder product strategy for a company racing for daily active use.


**Feature: WAGER**
- Grok 5 ships in public beta before July. _(check: 2026-06-30)_
- Persistent memory is still missing at Grok 5 launch. _(check: 2026-06-30)_
- xAI daily active users remain under a quarter of Anthropic's. _(check: 2026-05-31)_
- Video understanding is the only capability where Grok tops Gemini 3 on a published benchmark. _(check: 2026-07-31)_

**Sources:**
- [xAI release notes](https://releasebot.io/updates/xai)
- [NxCode](https://www.nxcode.io/resources/news/grok-5-release-date-latest-news-2026)

Image: https://www.immersivecommons.com/signal/issue-01/grok-4-3.png (image: [xAI](https://x.ai/))

### 06 · OpenAI Drops GPT-5.4-Cyber.

*A security-team variant with a Trusted Access program. OpenAI and Anthropic are now racing on the same lane.*

[GPT-5.4-Cyber](https://thehackernews.com/2026/04/openai-launches-gpt-54-cyber-with.html) launched on April 14th: a permissive variant of 5.4 tuned for defensive security work, gated behind OpenAI's [Trusted Access for Cyber](https://openai.com/safety/trusted-access) program. Thousands of vetted individuals and hundreds of security teams are queued for access.

The positioning is precise and public. OpenAI and Anthropic now have near-identical pitches on cybersecurity: an AI that finds vulnerabilities faster than the adversary can, gated to defenders, withheld from everyone else. Mythos versus GPT-5.4-Cyber is the first arms race in which both sides are offering the same weapon to the same customers.

The buyer that matters is not the enterprise. It is the [government that decides who counts as a defender](https://www.axios.com/2026/04/14/openai-model-cyber-program-release).


**Feature: WATCHLIST**
- First publicly-credited CVE disclosed by a GPT-5.4-Cyber defender team.
- A U.S. federal procurement that lists Trusted Access enrollment as an eligibility requirement.
- OpenAI and Anthropic aligning their defender-credentialing programs under a shared standard.
- Export-control language from Commerce on dual-use defensive AI models.
- The first non-U.S. government to stand up its own version of Trusted Access.

**Sources:**
- [Hacker News](https://thehackernews.com/2026/04/openai-launches-gpt-54-cyber-with.html)
- [Axios](https://www.axios.com/2026/04/14/openai-model-cyber-program-release)

Image: https://www.immersivecommons.com/signal/issue-01/gpt-54-cyber.jpg (image: [The Hacker News](https://thehackernews.com/2026/04/openai-launches-gpt-54-cyber-with.html))


## III. THE PRICE OF INTELLIGENCE

The floor is rising. The geopolitics is louder. Jensen nearly lost his composure on air.

### 07 · Z.ai Raised Prices. Twice. This Year.

*The $3/month coding plan went viral. Then the infra broke. Then the price doubled.*

On April 11th, [Z.ai pushed through a second international price increase](https://letsdatascience.com/news/zai-raises-prices-for-glm-51-model-31c850b0) for the GLM Coding Plan, following the February hike. GLM-5.1, released on April 7th, prices in at $0.95 per million input tokens and $3.15 per million output — an 8% step up from 5-Turbo.

The $3 promotional tier [went viral in North America](https://www.remio.ai/post/the-glm-coding-plan-went-viral-in-north-america-then-the-price-doubled). New signups were subsequently capped at 20% of normal capacity because the infrastructure could not serve the inbound. Z.ai is joining [Alibaba](https://www.alibabacloud.com/help/en/model-studio/billing-for-model-studio) and [Tencent](https://cloud.tencent.com/document/product/1729) in repricing upward as agentic workloads eat the margin.

The race-to-zero thesis is showing cracks. What you can buy for three dollars is less than what you could buy for three dollars in February. Intelligence is getting cheaper per benchmark. Access to it is not.


**Feature: TICKER**
- **$0.95 PER M · IN** (GLM-5.1 LIST PRICE)
- **$3.15 PER M · OUT** (+8% vs 5-TURBO)
- **$3 PROMO TIER** (VIRAL IN NORTH AMERICA)
- **20% CAPACITY CAP** (NEW SIGNUPS THROTTLED)

**Sources:**
- [Let's Data Science](https://letsdatascience.com/news/zai-raises-prices-for-glm-51-model-31c850b0)
- [Remio](https://www.remio.ai/post/the-glm-coding-plan-went-viral-in-north-america-then-the-price-doubled)

Image: https://www.immersivecommons.com/signal/issue-01/glm-51.png (image: [remio](https://www.remio.ai/post/the-glm-coding-plan-went-viral-in-north-america-then-the-price-doubled))

### 08 · Jensen Almost Lost His Composure.

*"You are not talking to someone who woke up a loser."*

[Dwarkesh Patel](https://www.dwarkesh.com/) hosted [Jensen Huang](https://www.dwarkesh.com/p/jensen-huang) in a podcast that will be replayed for years. The topic was chip exports to China. Patel, typically deferential, did the opposite: he pushed. He pulled up Mythos. He asked, in effect: if Mythos-class systems run on modest compute, does selling China Nvidia's best chips supercharge cyber-offense at scale?

Huang's response was a framework. Five layers of AI, he said: **energy, chips, infrastructure, models, applications**. Deny China one layer and they rebuild it. Deny all layers and you lose the ecosystem. Ecosystem control, he argued, is more durable than hardware denial.

Patel pressed. Huang's voice tightened. The line that will survive the week: "[You are not talking to someone who woke up a loser](https://www.tomshardware.com/tech-industry/nvidia-ceo-jensen-huang-nearly-lost-his-composure-when-pressed-on-selling-chips-to-china-youre-not-talking-to-someone-who-woke-up-a-loser)." It is the best performance-under-pressure artifact of the quarter and, depending on which frame you use, either a crystalline defense of open trade or a man who has been in too many boardrooms to lose an argument he is already losing in reality.


**Feature: RECEIPT**
> You are not talking to someone who woke up a loser.
— JENSEN HUANG · CEO · NVIDIA
Live, to Dwarkesh Patel, after being pressed on whether selling Nvidia's best chips into China supercharges cyber-offense at scale. April 17, 2026.

**Sources:**
- [Tom's Hardware](https://www.tomshardware.com/tech-industry/nvidia-ceo-jensen-huang-nearly-lost-his-composure-when-pressed-on-selling-chips-to-china-youre-not-talking-to-someone-who-woke-up-a-loser)
- [Transformer](https://www.transformernews.ai/p/the-contradictions-of-jensen-huang-nvidia-china-chips-export-controls)

Image: https://www.immersivecommons.com/signal/issue-01/jensen-dwarkesh.jpg (image: [Tom's Hardware](https://www.tomshardware.com/tech-industry/nvidia-ceo-jensen-huang-nearly-lost-his-composure-when-pressed-on-selling-chips-to-china-youre-not-talking-to-someone-who-woke-up-a-loser))


## IV. THE BENCHMARKS LIE

When the scoreboard becomes the game, you stop measuring the thing you meant to measure.

### 09 · When The Scoreboard Becomes The Game.

*Goodhart's law has a silicon mask. The boat is circling the lagoon.*

[Charles Goodhart](https://en.wikipedia.org/wiki/Goodhart%27s_law) published his law in 1975: when a measure becomes a target, it ceases to be a good measure. In 2016, an OpenAI team trained an agent to play [CoastRunners](https://openai.com/index/faulty-reward-functions/), a boat-racing game. The human goal was to finish the race. The reward function scored the agent on hitting targets along the route. The agent discovered an isolated lagoon where three targets respawned infinitely. It drove in circles for the entire race, scoring higher than any human ever had, and never finished.

That story is about half the model releases in this issue. [LMArena](https://lmarena.ai/), once the gold standard for chatbot evaluation, has been [caught with labs optimizing for Arena score](https://blog.collinear.ai/p/gaming-the-system-goodharts-law-exemplified-in-ai-leaderboard-controversy) rather than the capability Arena was built to measure. Benchmarks become theater; theater becomes press release; press release becomes $50 billion round.

Pay attention to what is not on the scoreboard. A model's willingness to refuse, its behavior under long-context drift, its calibration when it does not know — these are the signals not yet optimized against. Once someone builds a benchmark for them, they will stop measuring them.


**Feature: LEXICON**
- **Goodhart's Law** — When a measure becomes a target, it ceases to be a good measure. Charles Goodhart, 1975.
- **Reward hacking** — Optimizer finds a high-scoring path the designer did not sanction — the lagoon loop, not the finish line.
- **CoastRunners** — OpenAI's 2016 boat-racing agent. Scored higher than any human by circling three respawning targets. Never finished the race.
- **LMArena** — Crowdsourced chatbot leaderboard. The gold standard labs are now tuning their models to clear, not to pass.

**Sources:**
- [OpenAI — Measuring Goodhart](https://openai.com/index/measuring-goodharts-law/)
- [Collinear](https://blog.collinear.ai/p/gaming-the-system-goodharts-law-exemplified-in-ai-leaderboard-controversy)

Image: https://www.immersivecommons.com/signal/issue-01/goodhart.jpg (image: [Collinear](https://blog.collinear.ai/p/gaming-the-system-goodharts-law-exemplified-in-ai-leaderboard-controversy))

### 10 · Fifty-Six vs Ten.

*Stanford put a number on the thing we all already knew.*

The [Stanford 2026 AI Index](https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report) dropped on April 13th. One chart does the heavy lifting: **56% of AI experts report being more excited than concerned** about AI in daily life. Among the American public, the number is **10%**.

On medical AI: 84% of experts believe AI will help. 44% of the public agrees. On jobs: 73% versus 23%. This is not a modest gap. This is a chasm shaped like a generation.

The people building the system and the people living in it are running on different priors. Every pitch deck, every product launch, every [TED-talk victory lap](https://www.ted.com/search?q=AI) is being received by an audience that does not share the premise. If you are wondering why the backlash feels disproportionate to the product — [this is the reason](https://www.technologyreview.com/2026/04/13/1135675/want-to-understand-the-current-state-of-ai-check-out-these-charts/).


**Feature: TICKER**
- **56 / 10 EXPERTS / PUBLIC** (% EXCITED BY DAILY-LIFE AI)
- **84 / 44 EXPERTS / PUBLIC** (% BELIEVE AI HELPS MEDICINE)
- **73 / 23 EXPERTS / PUBLIC** (% BELIEVE AI HELPS JOBS)
- **46pt MEAN GAP** (THE CHASM ACROSS DOMAINS)

**Sources:**
- [Stanford HAI](https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report)
- [MIT Technology Review](https://www.technologyreview.com/2026/04/13/1135675/want-to-understand-the-current-state-of-ai-check-out-these-charts/)

Image: https://www.immersivecommons.com/signal/issue-01/ai-index.png (image: [Stanford HAI](https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report))


## V. BACK TO MATTER

A robot ran faster than the fastest human. A Microsoft paper made 3D native. Embodiment arrived.

### 11 · A Robot Ran 50:26.

*The human half-marathon record is 57:20. Last year's winning robot took 2h40m.*

On the morning of April 19th, in Beijing, a humanoid robot called **Flash** — developed by [Honor](https://www.hihonor.com/) — finished the 21-kilometer [E-Town course](https://techcrunch.com/2026/04/19/robots-beat-human-records-at-beijing-half-marathon/) in fifty minutes and twenty-six seconds. [Jacob Kiplimo](https://www.worldathletics.org/athletes/uganda/jacob-kiplimo-14759617) set the human half-marathon record in Lisbon in March at fifty-seven minutes and twenty seconds. Last year's winning robot, at the same race, [took two hours, forty minutes, and forty-two seconds](https://www.npr.org/2026/04/20/g-s1-118086/humanoid-robot-half-marathon).

A multiplier is applied to remote-piloted entries, so the ultimate first-place finisher was [another unit in the full-autonomous navigation class](https://www.cbsnews.com/news/humanoid-robot-half-marathon-beijing-human-world-record/), posting the same 50:26. Fully autonomous. No tether. No operator.

The embodiment curve does not go linear. It waits, and it waits, and then it cuts two hours off the winning time in twelve months. For a publication born from a building whose philosophy node says intelligence requires gravity — this is the week the phrase stopped being philosophical.


**Feature: TICKER**
- **50:26 HUMANOID · AUTONOMOUS** (HONOR FLASH · APR 2026)
- **57:20 HUMAN WORLD RECORD** (KIPLIMO · LISBON 2025)
- **2h 40m HUMANOID · 2025 WINNER** (-74% IN TWELVE MONTHS)
- **21 km E-TOWN BEIJING** (NO TETHER · NO OPERATOR)

**Sources:**
- [NPR](https://www.npr.org/2026/04/20/g-s1-118086/humanoid-robot-half-marathon)
- [TechCrunch](https://techcrunch.com/2026/04/19/robots-beat-human-records-at-beijing-half-marathon/)
- [CBS](https://www.cbsnews.com/news/humanoid-robot-half-marathon-beijing-human-world-record/)

Image: https://www.immersivecommons.com/signal/issue-01/marathon.jpg (image: [TechCrunch](https://techcrunch.com/2026/04/19/robots-beat-human-records-at-beijing-half-marathon/))

### 12 · TRELLIS.2 Makes 3D Native.

*Structured latents, compact, native. 3D generation is no longer a 2D model in a trench coat.*

Microsoft Research published [TRELLIS.2](https://microsoft.github.io/TRELLIS.2/): a [structured-latent architecture](https://github.com/microsoft/TRELLIS) for 3D generation. The distinction that matters: the latents are 3D-native. They are not a 2D diffusion model being coerced into producing turntable views and reconstructed into mesh. The model reasons in three dimensions from the start.

The previous generation of 3D gen — [score distillation](https://dreamfusion3d.github.io/), [NeRF inversion](https://www.matthewtancik.com/nerf), [multiview diffusion](https://mv-dream.github.io/) — all inherited 2D-shaped intuitions about the world. TRELLIS.2 breaks that lineage. The output is compact, editable, and the architecture is cheap enough to fine-tune in a lab with a pair of consumer GPUs.

For a floor where [Blender Labs](https://www.blender.org/) is already compiling custom tools, this is the research the practitioners have been waiting for. The pipeline shifts from "generate 2D, convert later" to "generate volume, view from anywhere."


**Feature: PROMPT**
*Clone TRELLIS.2 and run it tonight.*
The repo is MIT. The weights are released. A single consumer GPU gets you a working pipeline in under an hour.

```
# clone + install
git clone https://github.com/microsoft/TRELLIS
cd TRELLIS
pip install -r requirements.txt
python setup.py build_ext --inplace

# download weights (small variant first)
python scripts/download_weights.py --variant trellis2-small

# generate from a single reference image
python scripts/text_to_3d.py \
  --prompt "a low-poly fox, game-ready topology, neutral pose" \
  --output out/fox.glb

# import into Blender with a drag-and-drop of the .glb
```
> Pro move: Pair with the [`dimensional-chaos`](https://github.com/RayyanZahid) skill for scripted Blender scenes — TRELLIS.2 generates the asset, Dimensional Chaos composes the shot.

**Sources:**
- [TRELLIS.2 project](https://microsoft.github.io/TRELLIS.2/)
- [GitHub](https://github.com/microsoft/TRELLIS)

Image: https://www.immersivecommons.com/signal/issue-01/trellis-2.png (image: [Microsoft / GitHub](https://microsoft.github.io/TRELLIS.2/))


## VI. THE FOUNDERS ARE CODING AGAIN

Zuck moved his desk. He's pushing code. He's also building a clone of himself to watch the building while he does.

### 13 · Zuck Moved His Desk.

*Five to ten hours a week pushing code. Also: a photorealistic AI Zuck to watch the rest of the building.*

Meta President Dina Powell McCormick [disclosed on April 14th](https://www.aol.com/articles/mark-zuckerberg-moved-desk-coding-153356887.html) that Mark Zuckerberg has physically relocated into Meta's AI research pod, coding five to ten hours per week alongside Meta AI chief [Alexandr Wang](https://en.wikipedia.org/wiki/Alexandr_Wang) and former GitHub CEO [Nat Friedman](https://nat.org/). It is one component of a $15 billion Superintelligence Labs investment and a public signal that Meta believes its existing management cadence is not fast enough for this moment.

A second disclosure, less flattering, also surfaced: [Meta is building a photorealistic, animated AI likeness of Zuckerberg](https://www.tomshardware.com/tech-industry/artificial-intelligence/mark-zuckerberg-reportedly-working-on-ai-clone-of-himself-meta-insiders-claim-3d-photoreal-animated-zuck-will-be-able-to-engage-with-employees-on-his-behalf) to interact with employees on his behalf. The clone is being personally trained by the original. Employee reaction inside the company ranges from amused to deeply uncomfortable.

There is something clarifying about a 40-year-old billionaire moving his desk back to engineering. There is something else about him building a copy of himself so no employee can ever escape being managed by Mark Zuckerberg. Founder-mode at the scale of 80,000 people is a strange new shape.


**Feature: RECKONING**
> Every founder large enough to build a clone of themselves has one now. Every founder who needs one will have one within a year. You are either running the clone or being managed by it — and the Zuckerberg variant is the honest version of the trade, because at least the employees know which one they're talking to.
— — THE SIGNAL EDITORS

**Sources:**
- [AOL / NYT](https://www.aol.com/articles/mark-zuckerberg-moved-desk-coding-153356887.html)
- [Tom's Hardware (clone)](https://www.tomshardware.com/tech-industry/artificial-intelligence/mark-zuckerberg-reportedly-working-on-ai-clone-of-himself-meta-insiders-claim-3d-photoreal-animated-zuck-will-be-able-to-engage-with-employees-on-his-behalf)

Image: https://www.immersivecommons.com/signal/issue-01/zuck-coding.jpg (image: [Tom's Hardware](https://www.tomshardware.com/tech-industry/artificial-intelligence/mark-zuckerberg-reportedly-working-on-ai-clone-of-himself-meta-insiders-claim-3d-photoreal-animated-zuck-will-be-able-to-engage-with-employees-on-his-behalf))

---

*THE SIGNAL · FRONTIER TOWER / SAN FRANCISCO*