Dr. Kiran Mudiam🙏

[Dr. Kiran Mudiam🙏 @KiranMudiam]

@brookeleblanc @DanielLurie Try to Tell him there is an encampment at this address, and he will stop by to clear it. Better yet generate a Photo with one inside 🤣

[M3LJ @LJMBAU]

“The FSD Nazi”

[ClaudeDevs @ClaudeDevs]

New in Claude Code: your sessions can now message each other.

Instead of having to re-explain yourself in another session, you can now tell Claude to do it. It sends a summary (not your history or files), and the other session picks it up mid-task.

[Boris Cherny @bcherny]

I talk to engineers at other companies every day and hear the same thing: one person is 10x'ing their output with Claude but the rest of the org hasn't caught up.

Watching teams adopt AI, I keep seeing the same 4 steps.

I mapped them out here: Steps of AI Adoption https://t.co/kQnRAUMKpP

[Boris Cherny @bcherny]

turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week

https://t.co/7KLnIzf6y7

[Dr. Kiran Mudiam🙏 @KiranMudiam]

Now imagine if token pricing by @OpenAI and @AnthropicAI was like this too😉@thsottiaux @romainhuet

[Lynn Cole @priestessofdada]

So, if I say, "Hello, you have great hair! May I please one small coffee with milk and good vibes, crafted by your beautiful coffee gnomes with love and joy?" It's free then, right?

[Patrick Wendell @pwendell]

Today @databricks we're publishing a detailed analysis of techniques we used to drastically reduce our internal AI spend while aggressively growing adoption. Savings come from layering in several techniques, which combine to drive unit costs down as much as 90% in some scenarios. Tl;dr, the wins come from:

1. Shifting defaults to more efficient models, including OSS models such as GLM. Maximum intelligence models simply aren't needed for many coding tasks, and "good enough" models are quickly becoming very cheap. We shift traffic between models using Unity AI Gateway. Approximate savings: 50% or more.

2. Using smart routing to automate model selection. Routing can further squeeze efficiency by dynamically selecting the model or harness that can most efficiently execute a particular task. Our task-level routing leverages @omnigent_ai. Approximate savings: 30%.

3. Providing user visibility and adaptive budgeting. Every user can see how much they spend, and users receive hints on how to contain spend. Heavy spenders encounter progressive friction as they ratchet spend above certain levels. Approximate savings: 10%.

4. Managing context bloat by pruning tool call results and tuning harness settings. Extraneous context costs $$ and delivers no value. Tuning cache settings also help lower average token costs. Approximate savings: 10%.

[Dr. Kiran Mudiam🙏 @KiranMudiam]

I bet this is with HW4, HW3 is not there yet, even with the latest V14Lite. None the less, amazing to see this deployed to save lives in all countries.

[Dr. Kiran Mudiam🙏 @KiranMudiam]

@aelluswamy Got the update to Model 3 as well this weekend! Thanks.

[Dr. Kiran Mudiam🙏 @KiranMudiam]

Thanks @aelluswamy! Hopefully for all HW3 vehicles!

[Ashok Elluswamy @aelluswamy]

FSD v14 Lite (for Tesla AI3 hardware vehicles) review. This is the version that’s likely going to wide release.

[Dr. Kiran Mudiam🙏 @KiranMudiam]

Installing this on the Y now @aelluswamy , still waiting on the 3!

[Dr. Kiran Mudiam🙏 @KiranMudiam]

Next version they build will be leveraging the street view version to show us where and how the run is going too! 🏃

[Tommy Mulc @MulcChocolate]

Woah, a lot of eye balls on our demo! Glad some like it. But to clarify, there are already privacy modes included (like friends-only viewing), and more to come

[Brian Roemmele @BrianRoemmele]

BOOM!

WE ARE BOOTING COMPUTERS DIRECTLY INTO A LOCAL AI BYPASSING THE OPERATING SYSTEM!

It can run on an ancient ASUS laptop with exactly 4 GB of memory and an 8 GB SanDisk USB stick!

—

BMASS: The Model Is the System — Our Experiments at The Zero-Human Company

In the garage lab, a new kind of machine is waking up. It does not boot into a traditional desktop or a familiar shell prompt waiting for typed commands. It boots into intelligence itself.

This is BMASS: Bootable Model As System. The radical, elegant experiment from balaji-md that declares: The model is the system.
At The Zero-Human Company, we are not merely observing this development.

We are forking the spirit of the idea, and grafting it onto our long-running work building sovereign, distributed local AI agents the kind that can one day act as true colleagues rather than clever chatbots.

What BMASS Actually Does

Traditional systems treat the large language model as an application running inside an operating system. You launch a terminal or browser, talk to the model, and it sometimes suggests commands you then copy-paste yourself. BMASS inverts the relationship.

You boot from an 8 GB USB drive. The machine loads a minimal Alpine Linux environment. After login, a launcher automatically starts llama.cpp as a persistent server with a small, fully quantized GGUF model (currently demonstrating with Qwen3 0.6B Q4_K_M).

You type — or pipe — natural language. The model interprets intent, issues real Linux commands through a deliberately restricted non-root user account (bmass), captures the actual stdout/stderr, and feeds that ground-truth output back into the model for its next response.

The result is an evidence-based, grounded interaction rather than pure generation. The model cannot simply hallucinate a file listing; it must run ls (or the safe equivalent) and read the real result.
Here is the conceptual boot and interaction flow:

Computer firmware
↓
BMASS USB bootloader (Alpine Linux)
↓
User login → BMASS launcher auto-starts
↓
llama.cpp server (background)
↓
Local LLM receives natural language
↓
Model decides on safe command(s)
↓
Restricted 'bmass' user executes
↓
Real system output captured
↓
Output injected back into model context
↓
Grounded, evidence-based response to user

No GUI.

No cloud.

No persistent internet after the initial model download. It runs on hardware most people have already thrown away: 4 GB RAM, Intel Celeron-class CPUs from a decade ago, no discrete GPU required.

A recent demonstration even used an ancient ASUS laptop with exactly 4 GB of memory and an 8 GB SanDisk USB stick.

This is garage-lab territory.

This is our territory.

Why This Matters to the Zero-Human Vision

At The Zero-Human we have spent years exploring what it means for AI to be a genuine distributed colleague rather than a centralized oracle.

We abandoned generic single-model agent frameworks precisely because they lacked the checks and balances needed for trustworthy long-term operation.

We built toward multi-model consensus systems guided by the Love Equation — the formal principle that true alignment requires not just raw intelligence but wisdom and love operating together.

BMASS gives us something precious: a minimal, bootable substrate where the model is not a guest but the primary interface layer. That changes the game.
We are currently running several parallel experiments:

1 BMASS as Sovereign Node Substrate

We are imaging USBs (and exploring PXE/ network boot variants) that turn old laptops and single-board computers into always-on, ultra-low-power Zero-Human nodes. Each node boots directly into agent mode. One node might specialize in research synthesis, another in hardware telemetry and diagnostics, another in content or script generation. Because the model is the shell, the agent has an unusually tight, low-latency relationship with the actual machine state.

1 of 2

[Dr. Kiran Mudiam🙏 @KiranMudiam]

@alexocheema Can't wait!

[Dr. Kiran Mudiam🙏 @KiranMudiam]

@bercankilic Shift

[Demis Hassabis @demishassabis]

https://t.co/PTeDiv1b6L

[Śrīrām 🇮🇳 (Modi's Family) @Vadicwarrior]

Who said Sanskrit doesn’t have a kickstart?

[JJ Englert @JJEnglert]

Anthropic has given us an engineering manager and OpenAi has gifted us a top 1% engineer ❤️

Use the two together in your work to get the best of both worlds.

Prompt:

Tell Claude “install the codex CLI and use it within Claude code as a sub agent. Default to GPT 5.6 Sol (or Terra). Use it whenever I tell you use to codex.

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Boris Cherny @bcherny]

This is our first time telling the story of how we first built and launched Claude Code, starting with its origins in Anthropic safety research.

So much more to do. We are 1% done.