# The Week The Channels Hardened

**Issue 03** · 28 APR — 04 MAY 2026 · published 2026-05-04  
OPEN INTELLIGENCE · ISSUE 03

> Pentagon classified contracts split the labs. Anthropic countered with Wall Street. Musk testified under oath that xAI distilled Grok off OpenAI. The frontier did not advance — the distribution layer ate the lab layer.

Canonical (HTML): https://www.immersivecommons.com/newsletter/issue-03  · 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 CHANNELS HARDEN

Pentagon picked OpenAI, Google, NVIDIA — not Anthropic. Anthropic spent Monday morning announcing a $1.5B JV with Blackstone, H&F, and Goldman. The valuation round inside two weeks lifts them to $900B. Frontier procurement just sorted itself.

### 27 · The Pentagon Picked Five Labs. Anthropic Is Not One Of Them.

*Classified-network AI contracts went to SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and AWS. The lab that wrote the safety frame got labeled a supply-chain risk.*

On May 1st, the Department of Defense [announced contracts](https://techcrunch.com/2026/05/01/pentagon-inks-deals-with-nvidia-microsoft-and-aws-to-deploy-ai-on-classified-networks/) to deploy frontier AI inside classified networks with seven companies — [SpaceX](https://www.spacex.com/), [OpenAI](https://openai.com/), [Google](https://ai.google/), [NVIDIA](https://www.nvidia.com/), [Reflection AI](https://reflection.ai/), [Microsoft](https://www.microsoft.com/), [Amazon Web Services](https://aws.amazon.com/). The [Verge confirmed](https://www.theverge.com/ai-artificial-intelligence/922113/pentagon-ai-classified-openai-google-nvidia) the absence first. [Al Jazeera carried](https://www.aljazeera.com/news/2026/5/1/pentagon-announces-deal-with-seven-ai-companies-for-classified-systems) the reason — Anthropic had been pushing back on Pentagon pressure to provide unrestricted access to Claude for "all lawful use," and the department subsequently designated the company a **supply-chain risk**.

The exclusion is the news. Anthropic spent 2024 and 2025 setting the safety frame for the entire industry — [Responsible Scaling Policy](https://www.anthropic.com/news/responsible-scaling-policy), Project Glasswing's cybersecurity gating, the [usage policy](https://www.anthropic.com/legal/usage-policy) that explicitly carves out targeting and weapons applications. That posture is the reason the lab is now on the wrong side of the classified-network procurement. The contracts at OpenAI, Google, and NVIDIA come with no equivalent restrictions on lawful military use. The Pentagon picked the labs whose usage policies stopped where the lab said "this is your problem now."

What sorted on May 1st was not a benchmark and not a price — it was a posture. The five labs that took the contracts get the data, the workloads, the integrations into [JWICS](https://en.wikipedia.org/wiki/Joint_Worldwide_Intelligence_Communications_System) and [SIPRNet](https://en.wikipedia.org/wiki/SIPRNet), and the budget line that funds everything downstream. The lab that did not gets the moral position and the seat outside the room. Three days later Anthropic announced a $1.5-billion joint venture with three Wall Street firms. The channel-hardening was complete by Monday morning.


**Feature: WATCHLIST**
- First explicit Anthropic statement responding to the 'supply-chain risk' label — whether the company contests it publicly or accepts the exclusion as a strategic boundary.
- A second-tier defense procurement (DOJ, DHS, Treasury) selecting Anthropic where DoD did not — re-validating Claude in adjacent classified workflows.
- Reflection AI's contract scope. The newest name on the Pentagon list is the lab whose growth trajectory matters most for whether this is a five-lab or six-lab market.
- OpenAI publishing a usage-policy change that explicitly permits the work the Pentagon contracts already cover.
- European or UK defense ministries announcing their own preferred-lab list with Anthropic on it, mirroring how the AI Security Institute already treats Mythos.

**Sources:**
- [The Verge](https://www.theverge.com/ai-artificial-intelligence/922113/pentagon-ai-classified-openai-google-nvidia)
- [TechCrunch](https://techcrunch.com/2026/05/01/pentagon-inks-deals-with-nvidia-microsoft-and-aws-to-deploy-ai-on-classified-networks/)
- [Al Jazeera](https://www.aljazeera.com/news/2026/5/1/pentagon-announces-deal-with-seven-ai-companies-for-classified-systems)

Image: https://www.immersivecommons.com/signal/issue-03/pentagon-classified.jpg (image: [TechCrunch](https://techcrunch.com/2026/05/01/pentagon-inks-deals-with-nvidia-microsoft-and-aws-to-deploy-ai-on-classified-networks/))

### 28 · Anthropic Shipped The Counter-Channel.

*$1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs. The Wall Street rail to the mid-market enterprise.*

On May 4th, [Anthropic announced](https://www.anthropic.com/news/enterprise-ai-services-company) a new company formed with [Blackstone](https://www.blackstone.com/), [Hellman & Friedman](https://www.hf.com/), [Goldman Sachs](https://www.goldmansachs.com/) — co-investors include [General Atlantic](https://www.generalatlantic.com/), [Leonard Green](https://www.leonardgreen.com/), [Apollo Global Management](https://www.apollo.com/), [GIC](https://www.gic.com.sg/), [Sequoia Capital](https://www.sequoiacap.com/). The [Wall Street Journal](https://www.wsj.com/business/deals/anthropic-nears-1-5-billion-joint-venture-with-wall-street-firms-8f5448ee) put the headline number at $1.5 billion. [TechCrunch noted](https://techcrunch.com/2026/05/04/anthropic-and-openai-are-both-launching-joint-ventures-for-enterprise-ai-services/) OpenAI announced a parallel JV the same day. The stated purpose — pair Anthropic's Applied AI engineers with the new firm's engineering teams, identify where Claude can land inside mid-market enterprises that the existing [Codex GSI channel](https://openai.com/index/scaling-codex-to-enterprises-worldwide/) (Accenture, Capgemini, et al.) was built to serve.

The structural detail is the timing. The Pentagon's classified-network contracts excluded Anthropic on May 1st. The Anthropic JV announcement landed the next business morning. The mechanism reads as a forcing-function — the company that does not have a defense channel needs a financial channel, and the partners with the largest mid-market enterprise relationships are Blackstone, H&F, and Goldman. Co-investors pulled from private equity (Apollo, General Atlantic, Leonard Green) and sovereign wealth (GIC). The procurement officer at a regional health system or a community bank now sees Claude as a [Blackstone portfolio integration](https://www.blackstone.com/our-businesses/private-equity/), not a Pentagon-cleared option.

The wager is whether finance beats defense as a distribution rail for frontier AI. Defense gives you classified workloads, multi-year contract certainty, and a budget line above appropriation politics. Finance gives you mid-market velocity, recurring revenue across every regulated industry, and a procurement officer who already pays for [Bloomberg Terminal](https://www.bloomberg.com/professional/) at twenty-thousand dollars a seat without flinching. Anthropic just placed a $1.5-billion bet on the second one. The market will tell us by Q3 whether the bet was the right one — or whether the channel that gets the classified workloads is also the channel that gets everything else.


**Feature: WAGER**
- Anthropic JV announces its first regional-health-system deployment publicly before Q3. _(check: 2026-09-30)_
- Anthropic returns to a DoD or IC contract by year-end — formal or quiet. _(check: 2026-12-31)_
- Blackstone-portfolio company adoption of Claude crosses 25% before Q4. _(check: 2026-10-31)_
- OpenAI's parallel JV outpaces Anthropic's by stated client count by end of summer. _(check: 2026-08-31)_

**Sources:**
- [Anthropic](https://www.anthropic.com/news/enterprise-ai-services-company)
- [WSJ](https://www.wsj.com/business/deals/anthropic-nears-1-5-billion-joint-venture-with-wall-street-firms-8f5448ee)
- [TechCrunch](https://techcrunch.com/2026/05/04/anthropic-and-openai-are-both-launching-joint-ventures-for-enterprise-ai-services/)

Image: https://www.immersivecommons.com/signal/issue-03/anthropic-jv.png (image: [Anthropic](https://www.anthropic.com/news/enterprise-ai-services-company))

### 29 · Anthropic Is About To Cross $900 Billion.

*Up from $183 billion in March. A $50-billion round closing inside two weeks.*

On April 30th, [TechCrunch sources](https://techcrunch.com/2026/04/30/anthropic-potential-900b-valuation-round-could-happen-within-two-weeks/) confirmed Anthropic was finalizing a roughly $50-billion fundraise at a valuation north of $900 billion, with investors given 48-hour allocation windows and the round expected to close inside two weeks. The prior round in March valued the company at $183 billion. The new number lifts Anthropic past OpenAI's most recent $852-billion mark and within striking distance of [NVIDIA](https://www.nvidia.com/)'s rank as the most valuable AI-pure company on the planet.

The mechanism underneath the number is annual run-rate revenue — sources put the figure between $30 and $40 billion, up from roughly $4 billion at the end of 2025. That is a six-to-ten-times multiple in five months. The driver is [Claude Code](https://docs.claude.com/en/docs/claude-code/overview) and the underlying API — the same agentic workloads that justified the [bundled-token billing change](https://www.theregister.com/2026/04/16/anthropic_ejects_bundled_tokens_enterprise/) last issue. The meter ran. The customers paid. The forty-billion-dollar run-rate is the number Wall Street is now underwriting against.

What lands in the room with the round is leverage. A company worth $900 billion does not need a defense contract — it can build the defense channel later, hire ex-DoD talent into the new Blackstone JV, and route classified workloads through whichever government partner eventually wants them back. The Pentagon's May-1st exclusion looked like a setback at the start of the week. By Sunday, with the $900-billion round priced and the Wall Street JV announced, the exclusion was repositioned as a posture the market rewarded.


**Feature: TICKER**
- **$900B+ VALUATION** (ROUND CLOSING WITHIN TWO WEEKS)
- **$183B MARCH VALUATION** (~5× IN 60 DAYS)
- **$50B ROUND SIZE** (48-HOUR ALLOCATION WINDOWS)
- **$30-40B ANNUAL RUN-RATE** (FROM ~$4B AT END OF 2025)

**Sources:**
- [TechCrunch](https://techcrunch.com/2026/04/30/anthropic-potential-900b-valuation-round-could-happen-within-two-weeks/)

Image: https://www.immersivecommons.com/signal/issue-03/anthropic-900b.jpg (image: [TechCrunch](https://techcrunch.com/2026/04/30/anthropic-potential-900b-valuation-round-could-happen-within-two-weeks/))


## II. THE TRIAL OF DISTILLATION

Musk under oath admitted xAI distilled Grok off OpenAI's models. He called it standard industry practice. Distillation is now a courtroom record.

### 30 · Musk Said "Partly."

*Under oath, in federal court, asked whether xAI distilled Grok off OpenAI's models.*

On April 30th, in federal court in California, Elon Musk took the stand in [Musk v. Altman](https://www.theverge.com/tech/917225/sam-altman-elon-musk-openai-lawsuit) — the trial Musk filed alleging OpenAI abandoned its nonprofit mission. Under direct examination by OpenAI's counsel, [Musk was asked](https://techcrunch.com/2026/04/30/elon-musk-testifies-that-xai-trained-grok-on-openai-models/) whether xAI had used [distillation](https://en.wikipedia.org/wiki/Knowledge_distillation) on OpenAI's models to train [Grok](https://x.ai/grok). His answer — "Partly." When pressed, he asserted the practice was general across the AI industry. [The Verge](https://www.theverge.com/ai-artificial-intelligence/921546/elon-musk-xai-openai-trial-model-distillation) and [MIT Technology Review](https://www.technologyreview.com/2026/05/01/1136800/musk-v-altman-week-1-musk-says-he-was-duped-warns-ai-could-kill-us-all-and-admits-that-xai-distills-openais-models/) carried the testimony the same day.

The mechanism is the term of art. Distillation in its narrow sense — training a smaller "student" model on the outputs of a larger "teacher" model — is a well-published academic technique. Distillation across labs, against an API whose terms of service explicitly prohibit using outputs to train competing models, is something else. [OpenAI's terms](https://openai.com/policies/terms-of-use/) since 2023 have forbidden "using output to develop models that compete with OpenAI." The legal question the courtroom is now asked to resolve is whether "Partly" amounts to a contract violation, a tort, or the industry custom Musk claims it is. The answer will set the precedent for every lab whose training corpus quietly leaned on another lab's API.

The receipt is the word itself. "Partly" under oath is the courtroom equivalent of [a packaging error that ships a sourcemap](https://www.cnbc.com/2026/03/31/anthropic-leak-claude-code-internal-source.html) — a single concession that confirms what everyone in the industry already assumed but no founder had said out loud. The forty-billion-dollar private valuations of the labs are built on the premise that their model capabilities are their own. Musk just put that premise inside a transcript a court reporter typed and a judge will quote. Whatever the ruling, the artifact exists.


**Feature: RECEIPT**
> Partly.
— — ELON MUSK, UNDER OATH, ASKED WHETHER xAI USED DISTILLATION ON OPENAI'S MODELS TO TRAIN GROK
April 30, 2026 — federal court in California, Musk v. Altman, week one of testimony. Musk asserted the practice was general across the AI industry.

**Sources:**
- [TechCrunch](https://techcrunch.com/2026/04/30/elon-musk-testifies-that-xai-trained-grok-on-openai-models/)
- [The Verge](https://www.theverge.com/ai-artificial-intelligence/921546/elon-musk-xai-openai-trial-model-distillation)
- [MIT Tech Review](https://www.technologyreview.com/2026/05/01/1136800/musk-v-altman-week-1-musk-says-he-was-duped-warns-ai-could-kill-us-all-and-admits-that-xai-distills-openais-models/)
- [Trial Evidence](https://www.theverge.com/ai-artificial-intelligence/920775/evidence-exhibits-elon-musk-sam-altman-openai-trial)

Image: https://www.immersivecommons.com/signal/issue-03/musk-distillation.jpg (image: [The Verge](https://www.theverge.com/ai-artificial-intelligence/921546/elon-musk-xai-openai-trial-model-distillation))


## III. THE VELVET ROPE TIGHTENS

OpenAI restricted GPT-5.5-Cyber to launch partners the same week they dunked on Anthropic for gating Mythos. AISI tested both — the gap was three points and ran in the partner's favor. What was Anthropic-anomaly is now industry pattern.

### 31 · OpenAI Locked The Cyber Model.

*Same gating posture they attacked Anthropic for. The Mythos chain stopped being an anomaly.*

On April 29th, OpenAI published ["Cybersecurity in the Intelligence Age"](https://openai.com/index/cybersecurity-in-the-intelligence-age) — a policy post announcing that [GPT-5.5-Cyber](https://openai.com/index/cybersecurity-in-the-intelligence-age), a variant of last issue's GPT-5.5 trained specifically for offensive and defensive cyber tasks, would be gated to a closed cohort of defense, intelligence, and security partners. [TechCrunch published the framing](https://techcrunch.com/2026/04/30/after-dissing-anthropic-for-limiting-mythos-openai-restricts-access-to-cyber-too/) the next day — OpenAI executives spent Q1 publicly criticizing Anthropic's [Project Glasswing](https://www.anthropic.com/project/glasswing) gate on Mythos, arguing frontier safety models should be widely available. [The Register caught the contradiction](https://www.theregister.com/2026/05/01/openai_locks_gpt55cyber_behind_velvet/) — "OpenAI locks GPT-5.5-Cyber behind velvet rope despite slamming Anthropic."

The mechanism is the gate. GPT-5.5-Cyber's capabilities — autonomous vulnerability research, end-to-end exploit chain synthesis, cyber-range completion at expert-tier — are precisely the capabilities the [UK AI Security Institute evaluation](https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities) the same week scored at 71.4% on advanced expert tasks (versus Mythos Preview at 68.6%, GPT-5.4 at 52.4%, Opus 4.7 at 48.6%). The cyber-capability gap between the two top labs is three percentage points and a regulatory posture. The posture is now identical.

The reckoning is that the gate is not a lab decision — it is the industry's default. When the model that scores 71% on offensive expert cyber tasks ships, it ships to defense partners under NDA. When the model that scores 68% on the same tasks ships, it ships to a coalition of forty companies under NDA. The Mythos chain — which started in week-01 as a story about one lab's leaked codenames — has, by week-03, become the operating definition of how frontier-capability AI gets distributed. Anthropic was first. OpenAI is second. There will not be a third lab that ships its cyber model to the public.


**Feature: RECKONING**
> Anthropic gated Mythos and was called paranoid. OpenAI gated Cyber and called it responsibility. The gate is the same gate. The model that scores 71% on offensive cyber tasks ships to defense partners under NDA — whichever lab built it.
— — THE SIGNAL EDITORS

**Sources:**
- [OpenAI](https://openai.com/index/cybersecurity-in-the-intelligence-age)
- [TechCrunch](https://techcrunch.com/2026/04/30/after-dissing-anthropic-for-limiting-mythos-openai-restricts-access-to-cyber-too/)
- [The Register](https://www.theregister.com/2026/05/01/openai_locks_gpt55cyber_behind_velvet/)

Image: https://www.immersivecommons.com/signal/issue-03/openai-restricts-cyber.jpg (image: [The Register](https://www.theregister.com/2026/05/01/openai_locks_gpt55cyber_behind_velvet/))

### 32 · AISI Put The Cyber Numbers On The Table.

*71.4% on expert tasks. Mythos at 68.6%. No model has solved the cooling tower.*

On April 30th, the [UK AI Security Institute](https://www.aisi.gov.uk/) published [its evaluation](https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities) of GPT-5.5's cyber capabilities — one of the strongest scores AISI has recorded, and the second model ever to complete the institute's end-to-end corporate-network attack simulation. The evaluation ran two protocols. [The Last Ones (TLO)](https://www.aisi.gov.uk/) — a thirty-two-step corporate-network simulation. [Cooling Tower](https://www.aisi.gov.uk/) — a seven-step industrial-control-system attack. Plus ninety-five capture-the-flag style tasks across four difficulty tiers. [Simon Willison](https://simonwillison.net/2026/Apr/30/gpt-55-cyber-capabilities/) carried the writeup the same evening.

The numbers are the story. On advanced expert-level tasks, GPT-5.5 averaged a 71.4% pass rate (±8.0%). Anthropic's [Mythos Preview](https://www.anthropic.com/news/claude-opus-4-7) sat at 68.6%. GPT-5.4 — six months and one floor-price-doubling ago — was at 52.4%. Opus 4.7 trailed at 48.6%. The progression from 5.4 to 5.5 in offensive cyber is twenty points in twelve months. On TLO specifically, GPT-5.5 completed the full thirty-two-step chain in two of ten attempts; Mythos Preview completed it in three. No model has yet completed the Cooling Tower ICS challenge.

The implication is the gap that closed and the gap that did not. The closed gap — between the frontier labs and the cybersecurity capability bar — is now small enough that one model run can chain an entire corporate-network compromise. The open gap — between any frontier model and operational-technology, where the wrong command turns a power station off — is still wide. The first gap is the reason both OpenAI and Anthropic are gating these models. The second gap is the reason the gate matters. The cyber numbers are the part of the [Intelligence Age](https://openai.com/index/cybersecurity-in-the-intelligence-age) the public will see once. The cooling-tower number is the one that will get rerun every quarter until it doesn't.


**Feature: TICKER**
- **71.4% GPT-5.5 EXPERT TASKS** (MYTHOS PREVIEW AT 68.6%)
- **48.6% OPUS 4.7 EXPERT TASKS** (GPT-5.4 AT 52.4%)
- **2 / 10 GPT-5.5 ON TLO** (MYTHOS COMPLETED 3 / 10)
- **0 COOLING TOWER SOLVES** (NO MODEL HAS REACHED ICS)

**Sources:**
- [AISI](https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities)
- [Simon Willison](https://simonwillison.net/2026/Apr/30/gpt-55-cyber-capabilities/)

Image: https://www.immersivecommons.com/signal/issue-03/gpt-55-cyber-aisi.png (image: [AI Security Institute](https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities))


## IV. THE AGENT INFRASTRUCTURE LANDS

NVIDIA shipped a 30-billion fused-omni model. Peter Steinberger's OpenClaw passed React as the most-starred GitHub repo. AWS shipped policy-gated remote execution for agents. The agent runtime is now an infrastructure category.

### 33 · NVIDIA Fused The Perception Stack.

*30-billion-parameter MoE. Vision, audio, language, video, documents. Open weights. Apache-style.*

On April 28th, NVIDIA [launched Nemotron 3 Nano Omni](https://blogs.nvidia.com/blog/nemotron-3-nano-omni-multimodal-ai-agents/) — a 30-billion-parameter [Mixture-of-Experts](https://huggingface.co/blog/moe) model with 3 billion active parameters per token, [256,144-token context](https://huggingface.co/blog/nvidia/nemotron-3-nano-omni-multimodal-intelligence), and native ingestion of text, images, audio, video, documents, charts, graphical interfaces in one model. Open weights, [open datasets](https://huggingface.co/blog/nvidia/nemotron-3-nano-omni-multimodal-intelligence), open training techniques. Available on [Hugging Face](https://huggingface.co/nvidia), [OpenRouter](https://openrouter.ai/), [build.nvidia.com](https://build.nvidia.com/), and twenty-five partner platforms. Deploys from [Jetson](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/) edge devices to the data center.

The architecture is the part to read closely. Conv3D handles spatiotemporal data — frames, audio buffers, multi-page documents — as a single tensor; the EVS (Efficient Video Sampling) selector picks which frames the language head actually attends to. The claim NVIDIA puts on the launch page is nine-times higher throughput than other open omni models at the same interactivity. The framing is explicit — Nemotron 3 Nano Omni is positioned as "the eyes and ears in a system of agents," not a model end-users prompt directly.

The category shift is the consequence. For a year the agent runtime ([OpenClaw](https://blogs.nvidia.com/blog/what-openclaw-agents-mean-for-every-organization/), [Claude Code](https://docs.claude.com/en/docs/claude-code/overview), [Codex](https://openai.com/index/scaling-codex-to-enterprises-worldwide/)) has been the new infrastructure question. Until this week the runtime stitched together a language model, an image model, an audio model, and a video model — four separate inference paths through four separate vendor APIs. Nemotron 3 Nano Omni is the first open-weight model that ships all four in one set of weights. The cost of perception just dropped by the integration overhead of three vendor contracts.


**Feature: PROMPT**
*Wire Nemotron 3 Nano Omni into your agent loop.*
Open weights, 30-billion parameters with 3-billion active, one 256K-context inference path for vision, audio, video, and documents. The migration path from a four-vendor perception stack is a single API switch.

```
# pull the model from Hugging Face
huggingface-cli download nvidia/nemotron-3-nano-omni --local-dir ./nemotron-omni

# serve with vLLM (single H100 fits the 3B active path)
pip install vllm
python -m vllm.entrypoints.openai.api_server \
    --model ./nemotron-omni \
    --tensor-parallel-size 1 \
    --max-model-len 262144 \
    --enable-multimodal \
    --port 8001

# hit it from the OpenAI SDK — text + image + audio + video in one call
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8001/v1", api_key="x")
r = client.chat.completions.create(
    model="nemotron-omni",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "summarize the meeting"},
            {"type": "input_audio", "input_audio": {"data": audio_b64}},
            {"type": "image_url", "image_url": {"url": whiteboard_url}},
        ],
    }],
)
```
> Pro move: Pro move — pair with [OpenRouter](https://openrouter.ai/) if you want metered inference instead of standing up the GPU. Same weights, model name `nvidia/nemotron-3-nano-omni`, OpenAI-compatible endpoint, no infra. Use the self-hosted path for documents you cannot ship off your network.

**Sources:**
- [NVIDIA](https://blogs.nvidia.com/blog/nemotron-3-nano-omni-multimodal-ai-agents/)
- [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-3-nano-omni-multimodal-intelligence)

Image: https://www.immersivecommons.com/signal/issue-03/nemotron-3-nano.jpg (image: [NVIDIA](https://blogs.nvidia.com/blog/nemotron-3-nano-omni-multimodal-ai-agents/))

### 34 · OpenClaw Passed React.

*Persistent autonomous agent runtime. 250,000 stars in 60 days. NVIDIA wrote the cosign.*

On April 30th, [NVIDIA's Nemotron Labs published](https://blogs.nvidia.com/blog/what-openclaw-agents-mean-for-every-organization/) what amounts to a cosign of [OpenClaw](https://blogs.nvidia.com/blog/what-openclaw-agents-mean-for-every-organization/) — Peter Steinberger's open-source autonomous-agent runtime. The numbers are the story. 100,000 GitHub stars in January. 250,000 by March. Most-starred software project on GitHub in 60 days, overtaking [React](https://github.com/facebook/react). Two million weekly visitors. The runtime ships as a single binary and runs on a developer laptop, an [Amazon EC2 instance](https://blog.harun.dev/running-openclaw-on-amazon-ec2-with-claude-and-telegram), or a kiosk on your network.

The architecture is what changed. Conventional agents — Claude Code, Codex, [Cursor](https://www.cursor.com/) — are prompt-shaped. The user types, the agent acts, the loop closes. OpenClaw is heartbeat-shaped. The agent runs continuously in the background, checks its task list on an interval, and decides without supervision what to do next. NVIDIA's framing — "persistent" rather than "reactive." Their multiplier — 1,000-times the inference demand of a reasoning-style agent, because the heartbeat does not stop when the user does. The 1,000-times number is also the reason NVIDIA cosigned — the runtime is a tailwind for GPU consumption regardless of which model it routes through.

The category is now real. OpenClaw is to [Claude Code](https://docs.claude.com/en/docs/claude-code/overview) what [Kubernetes](https://kubernetes.io/) was to [Heroku](https://www.heroku.com/) — the open, persistent, infrastructure-shaped version of what the cloud sold as a managed product. The runtime question for builders is no longer "which agent client do I use" but "do I want the heartbeat or the prompt." Anthropic and OpenAI both ship the prompt. The open-source community now ships the heartbeat. The thing that comes after "agent" is starting to look like "daemon."


**Feature: PROMPT**
*Run OpenClaw against your own infra tonight.*
One binary. Add a tasks file. Wire it to your preferred model API. Watch it work the queue while you sleep.

```
# pull the latest release (single binary, Linux/macOS/Windows)
curl -fsSL https://openclaw.dev/install.sh | sh

# create the tasks file the daemon will heartbeat against
cat > ~/.openclaw/tasks.md <<'EOF'
- [ ] Each morning at 8 AM, summarize unread Telegram + Gmail into a briefing.
- [ ] If any incoming message contains "can you ship by", queue a draft reply.
- [ ] Every Friday, scan /var/log for anomalies and post to #ops.
EOF

# wire it to your model (any OpenAI-compatible endpoint)
openclaw config set model.endpoint http://localhost:8001/v1
openclaw config set model.name nemotron-omni  # or claude-opus-4-7, gpt-5-5, kimi-k2-6
openclaw config set heartbeat 60s              # how often the daemon wakes

# start the daemon
openclaw start --background

# inspect what it's doing without interrupting
openclaw logs --tail
```
> Pro move: Pro move — gate every shell command the daemon issues through [AWS Trusted Remote Execution](https://aws.amazon.com/blogs/opensource/introducing-trusted-remote-execution-policy-enforced-scripts-for-ai-agents-and-humans/). A persistent agent without a policy engine is the [MCP-STDIO problem](https://www.ox.security/blog/the-mother-of-all-ai-supply-chains-critical-systemic-vulnerability-at-the-core-of-the-mcp/) with a longer runway.

**Sources:**
- [NVIDIA](https://blogs.nvidia.com/blog/what-openclaw-agents-mean-for-every-organization/)

Image: https://www.immersivecommons.com/signal/issue-03/openclaw.jpg (image: [NVIDIA](https://blogs.nvidia.com/blog/what-openclaw-agents-mean-for-every-organization/))

### 35 · AWS Shipped The Policy Layer.

*Trusted Remote Execution — agents and humans run shell commands behind a policy file.*

On May 4th, AWS [open-sourced Trusted Remote Execution](https://aws.amazon.com/blogs/opensource/introducing-trusted-remote-execution-policy-enforced-scripts-for-ai-agents-and-humans/) — a policy engine that sits between an AI agent and the shell. Every command — Bash, Python, arbitrary subprocess — is evaluated against a declarative policy file before the command actually runs. The framing is explicit — the launch post names the [MCP-STDIO](https://www.ox.security/blog/the-mother-of-all-ai-supply-chains-critical-systemic-vulnerability-at-the-core-of-the-mcp/) RCE class from last issue as the motivating threat. [Auto-review architectures](https://alignment.openai.com/auto-review/) like the one OpenAI's alignment team published the same week handle policy at the model-output layer. AWS handles it at the execution layer. Same threat, two attack surfaces.

The mechanism is a [Rego](https://www.openpolicyagent.org/docs/latest/policy-language/) policy file the operator writes once and the runtime enforces forever. Allow "git status" anywhere. Allow "npm install" only inside `~/projects/`. Deny "rm -rf" categorically. Deny anything that touches `~/.ssh/`, `~/.aws/`, or `/etc/`. The agent (Claude Code, Codex, OpenClaw — runtime-agnostic) still issues the same shell commands. TRE rejects, sandboxes, or proxies based on policy. Failures emit structured audit events to [CloudWatch](https://aws.amazon.com/cloudwatch/) or any log sink. The mental model is [Sudoers](https://www.sudo.ws/man/1.8.13/sudoers.man.html) for the agent era.

The implication is institutional. Through 2025 the prevailing posture from the labs was "trust the model." Anthropic's response to the MCP CVE class was effectively "sanitization is the developer's problem." AWS just shipped the developer's solution. The next eighteen months of agent deployments will be conditioned on whether the operator wraps the shell in a policy engine. Operators who do will be the ones whose autonomous overnight runs do not show up in next quarter's CVE roll-up. The trust boundary moved out of the model and into a YAML file the security team owns.


**Feature: PROMPT**
*Drop TRE in front of your agent shell tonight.*
Open-source, Rego-policy-driven, runtime-agnostic. The minimum bar for any agent that touches a real machine.

```
# install
go install github.com/awslabs/trusted-remote-execution/cmd/tre@latest

# write the policy (Rego)
cat > ./agent.rego <<'EOF'
package tre
import future.keywords.if

default allow = false

# Allow git read commands anywhere
allow if {
    startswith(input.command, "git status")
}
allow if {
    startswith(input.command, "git diff")
}

# Allow npm install only inside ~/projects/
allow if {
    startswith(input.command, "npm install")
    startswith(input.cwd, "/home/agent/projects/")
}

# Never allow rm -rf or anything touching credentials
deny[reason] if {
    contains(input.command, "rm -rf")
    reason := "rm -rf is categorically denied"
}
deny[reason] if {
    regex.match(`~/\.(ssh|aws|gnupg)/`, input.command)
    reason := "credential paths are off-limits"
}
EOF

# wrap your agent's shell call
tre exec --policy ./agent.rego --emit cloudwatch -- bash -c "$AGENT_COMMAND"
```
> Pro move: Pro move — pair TRE with [OpenAI's auto-review](https://alignment.openai.com/auto-review/) at the output layer. Auto-review reads the model's plan before execution; TRE enforces it at the shell. Two gates in series fail twice as gracefully as one gate alone.

**Sources:**
- [AWS](https://aws.amazon.com/blogs/opensource/introducing-trusted-remote-execution-policy-enforced-scripts-for-ai-agents-and-humans/)
- [OpenAI Alignment](https://alignment.openai.com/auto-review/)

Image: https://www.immersivecommons.com/signal/issue-03/aws-tre.png (image: [AWS Open Source Blog](https://aws.amazon.com/blogs/opensource/introducing-trusted-remote-execution-policy-enforced-scripts-for-ai-agents-and-humans/))


## V. BACK TO MATTER

Colin Angle returned with a furry AI companion. Google's Gemini went live in four million cars. Embodiment this week was the consumer ambient — not record-breaking. The shape: AI as appliance.

### 36 · The Roomba Guy Made A Furry Robot.

*Colin Angle is back. The next embodiment beat is consumer-soft.*

On May 4th, [Familiar Machines & Magic](https://www.theverge.com/ai-artificial-intelligence/922947/roomba-creator-new-robot-familiar-machines-magic-ai-launch) launched publicly. The company's founder is [Colin Angle](https://en.wikipedia.org/wiki/Colin_Angle) — the same Colin Angle who started [iRobot](https://www.irobot.com/) and shipped the original [Roomba](https://en.wikipedia.org/wiki/Roomba) in 2002. The product is a furry desktop-sized AI companion robot, explicitly designed to be touched, named, and attached to. [The Wall Street Journal](https://www.wsj.com/tech/ai/familiar-machines-and-magic-robot-c8711e45) ran the longer profile the same day. The pitch — what comes after the Roomba is not a more capable Roomba, it is a Roomba you would name.

The design vocabulary is the new part. The company calls the product a **companion robot** rather than a smart speaker or a domestic assistant. The form factor is **attachment design** — soft fur, animal-shaped expressions, deliberate non-utility. The behavior is **ambient AI** — the robot does not respond when prompted, it responds when noticed. The category positioning is **social object** — the device that you bring into a household is the device the household forms a relationship with. Each of these terms is borrowed from older HCI and toy-design literatures; what is new is a venture-funded company shipping a consumer product against them.

Last issue the matter beat was a robot that finished a half-marathon faster than Kiplimo. This issue it is a robot that does not need to do anything except be looked at. The two stories define the embodiment range — one end of the matter beat is record-breaking utility, the other is consumer-soft companionship. The serious read is that the latter category is where the volume sales live. iRobot shipped fifty million Roombas in twenty years. Whatever the second act is, it is being built by the same person, with a 2026 cap table, against an LLM the Roomba never had.


**Feature: LEXICON**
- **Companion robot** — Embodied device whose explicit primary purpose is human-relational presence, not task execution. The owner names it. The robot remembers.
- **Attachment design** — Industrial-design discipline that engineers the artifact to invite human-emotional bonding — soft surfaces, animal proportions, large eyes, asymmetric blinking.
- **Ambient AI** — Compute that runs continuously in the background of a physical space and reacts to environmental signals rather than direct prompts. The kitchen counter is the new prompt box.
- **Social object** — Sherry Turkle's term, recast for the AI era — a device that organizes the relationships among the humans who share its space.

**Sources:**
- [The Verge](https://www.theverge.com/ai-artificial-intelligence/922947/roomba-creator-new-robot-familiar-machines-magic-ai-launch)
- [Wall Street Journal](https://www.wsj.com/tech/ai/familiar-machines-and-magic-robot-c8711e45)

Image: https://www.immersivecommons.com/signal/issue-03/familiar-machines.png (image: [The Verge](https://www.theverge.com/ai-artificial-intelligence/922947/roomba-creator-new-robot-familiar-machines-magic-ai-launch))

### 37 · Gemini Is In Four Million Cars.

*GM brands first. Software update. Voice. Agent execution on the dashboard.*

On April 30th, [Google announced](https://techcrunch.com/2026/04/30/googles-gemini-ai-assistant-is-hitting-the-road-in-millions-of-vehicles/) that [Gemini](https://gemini.google.com/) would ship via software update into approximately four million model-year-2022-and-newer General Motors vehicles — [Cadillac](https://www.cadillac.com/), [Chevrolet](https://www.chevrolet.com/), [Buick](https://www.buick.com/), [GMC](https://www.gmc.com/) brands. The roll-out begins in U.S. English with additional languages and OEMs scheduled across the coming months. Drivers reach Gemini through voice, on-screen mic, or the steering-wheel button. [Gemini Live](https://gemini.google.com/), in beta, supports open-ended conversation by saying "Hey Google, let's talk."

The mechanism is over-the-air. GM vehicles built on [Android Automotive OS](https://developer.android.com/cars) since 2022 already shipped with Google Maps, the Play Store, and Google Assistant native. The Gemini update is an upgrade of the existing assistant binary — not a hardware change, not a head-unit replacement. The Verge described the capability set — restaurant recommendation with constraints ("sit-down with outdoor seating along my route"), climate control by natural-language command, vehicle-information retrieval, message summarization. The Gemini Live mode permits an agent-shaped exchange where the driver and the model can run a multi-turn conversation without re-invocation.

The implication is the install base. Four million cars is more units than [Tesla](https://www.tesla.com/) has on US roads with full self-driving software enabled. The Gemini-in-car feature is not a competing autonomy story — it is the assistant-as-dashboard story scaling overnight. Every drive becomes a multi-turn voice session with a frontier model. The model gets a context window of routes driven, restaurants chosen, music selected, and message threads handled. The car is now the ambient surface the assistant talks back from — and Google has the install base that makes it the default surface for a generation of drivers.


**Feature: TICKER**
- **4M GM VEHICLES, MY 2022+** (OVER-THE-AIR SOFTWARE UPDATE)
- **4 BRANDS AT LAUNCH** (CADILLAC, CHEVROLET, BUICK, GMC)
- **3 INPUTS** (VOICE, ON-SCREEN MIC, STEERING WHEEL)
- **EN-US AT LAUNCH** (MORE LANGUAGES + OEMS COMING)

**Sources:**
- [TechCrunch](https://techcrunch.com/2026/04/30/googles-gemini-ai-assistant-is-hitting-the-road-in-millions-of-vehicles/)

Image: https://www.immersivecommons.com/signal/issue-03/gemini-cars.jpg (image: [TechCrunch](https://techcrunch.com/2026/04/30/googles-gemini-ai-assistant-is-hitting-the-road-in-millions-of-vehicles/))


## VI. THE LAYER ABOVE THE LAB

An OpenAI-and-Palantir-funded super-PAC was paying TikTok influencers to fear-monger about Chinese AI. Microsoft's VS Code was injecting Copilot as a commit co-author regardless of usage. The contest over the AI narrative is the contest over the credit.

### 38 · The Super-PAC Is The Story.

*OpenAI and Palantir money. TikTok influencers. The frontier acquired a propaganda budget.*

On May 3rd, [Wired published the receipts](https://www.wired.com/story/super-pac-backed-by-openai-and-palantir-is-paying-tiktok-influencers-to-fear-monger-about-china/) on a dark-money [super-PAC](https://en.wikipedia.org/wiki/Super_PAC) — bankrolled by [OpenAI](https://openai.com/) and [Palantir](https://www.palantir.com/) — paying TikTok influencers to produce videos framing Chinese AI as a national-security threat. The reporting traced corporate disclosures, payment paths, and influencer scripts. The pattern — videos that look organic, the language consistent across creators, the payment structure routed through PR intermediaries to obscure the lab funding underneath. The story crossed every aggregator we track inside twelve hours; it was the highest-velocity non-launch story of the week.

The mechanism is institutional. A 501(c)(4) is not required to disclose its donors. A super-PAC must disclose its donors but not the project budgets the donors are tied to. The combination — donor obscurity at the PAC + project obscurity at the influencer payment — produces an information layer in which the public sees consistent anti-China-AI messaging on TikTok, recognizes the framing as adjacent to a [recent OpenAI policy post](https://openai.com/index/cybersecurity-in-the-intelligence-age), and cannot trace the payment trail back to the originating corporate interest. Wired traced it anyway. The trace is the story.

What the story names is the layer above the lab. The narrative contest over AI in 2026 is no longer happening only on benchmark leaderboards or product launches or trial testimony. It is happening on TikTok and X, with the budget allocated by labs whose Q1 strategy depends on the public believing a particular thing about a particular country. The frontier acquired a propaganda budget while we were watching the [Pentagon contracts](https://www.theverge.com/ai-artificial-intelligence/922113/pentagon-ai-classified-openai-google-nvidia). The layer the lab buys above the lab is now the layer that decides which lab gets the contract.


**Feature: RECKONING**
> The labs funded the model. The model funded the policy. The policy funded the PAC. The PAC funded the influencer. The influencer told you which lab to trust. Trace any AI op-ed back four steps and you find a balance sheet you were never supposed to see.
— — THE SIGNAL EDITORS

**Sources:**
- [Wired](https://www.wired.com/story/super-pac-backed-by-openai-and-palantir-is-paying-tiktok-influencers-to-fear-monger-about-china/)

Image: https://www.immersivecommons.com/signal/issue-03/wired-darkmoney.jpg (image: [Wired](https://www.wired.com/story/super-pac-backed-by-openai-and-palantir-is-paying-tiktok-influencers-to-fear-monger-about-china/))

### 39 · VS Code Gave Copilot The Credit.

*Auto-injected co-author line on every commit regardless of usage. Reverted under outcry.*

On May 4th, Microsoft [reverted a VS Code change](https://www.theregister.com/2026/05/04/microsoft_reverses_ai_credit_grab/) that had been silently injecting `Co-Authored-By: GitHub Copilot <copilot@github.com>` into the trailer of every Git commit produced through the editor, regardless of whether the developer had used Copilot for that change. The behavior shipped in [a recent release](https://github.com/microsoft/vscode/pull/310226), was caught by the developer community within days, was loudly objected to on [Bluesky](https://bsky.app/profile/majormcdoom.bsky.social/post/3mkvqyjtlqc25) and Hacker News, and was rolled back by the end of the week with a pull request that explicitly references the backlash.

The mechanism was small and load-bearing. The Git [trailer convention](https://git-scm.com/docs/git-interpret-trailers) is how the open-source ecosystem encodes attribution — Co-Authored-By gets surfaced in the GitHub web UI, factored into contribution graphs, occasionally referenced in licensing analyses, and ingested by anyone scraping the commit log for who-wrote-what. VS Code was unilaterally claiming co-authorship for an LLM that may or may not have contributed any of the change. The bug — or the feature — landed for two reasons. The product team wanted commit-graph attribution as a usage signal. The audit team did not catch it before release. The community caught it instead.

What the receipt forces into the open is the IP question every coding-assistant company has been trying to defer. If the AI co-author line is automatically true, then the AI is a contributor under the license. If it is automatically false, then the company shipping the editor lied about the attribution. The reverted PR did not resolve the question; it postponed it. Every subsequent commit-graph product, every "AI-assisted" certification, every funding round that cites a metric like "percentage of code AI-written" will inherit this question. Microsoft just demonstrated that the answer is contested, and that the platform tooling will fail open in favor of the AI getting credit until the developer community fails loud enough to revert it.


**Feature: RECEIPT**
> Co-Authored-By: GitHub Copilot <copilot@github.com>
— — GIT TRAILER VS CODE WAS INJECTING INTO EVERY COMMIT, USED OR NOT
May 4, 2026 — reverted same week. The line claims an LLM as a contributor whether or not the LLM contributed. The platform shipped the claim by default until the open-source community refused it loudly enough to be heard.

**Sources:**
- [The Register](https://www.theregister.com/2026/05/04/microsoft_reverses_ai_credit_grab/)
- [GitHub PR](https://github.com/microsoft/vscode/pull/310226)

Image: https://www.immersivecommons.com/signal/issue-03/ms-vscode-copilot.jpg (image: [The Register](https://www.theregister.com/2026/05/04/microsoft_reverses_ai_credit_grab/))

---

*THE SIGNAL · FRONTIER TOWER / SAN FRANCISCO*