Hunee

Ninja. Regular geek...

[Sam Witteveen @Sam_Witteveen]

Full breakdown on YT 👇

https://t.co/rxuEibLIsa

[Sam Witteveen @Sam_Witteveen]

Most agent frameworks put safety rules in the system prompt and hope the model follows them. Wrong layer — one injection and they're gone.

OpenShell enforces policy via a YAML file.

A supervisor sits outside the agent and controls:
🌐 network — default deny
📁 filesystem — sandboxed workspace
🧠 inference — managed endpoint
🔑 credentials — injected at runtime

Compromise the agent, the policies still hold.

[Sam Witteveen @Sam_Witteveen]

OpenShell is key for Agents - NemoClaw is a blueprint:

- Harness (swappable)
- Model (swappable)
- Runtime — always OpenShell

The runtime is the part doing the security work. Swap everything else, that stays. That's the piece worth understanding.

DeepAgents can be that harness!

@NVIDIAAI @NVIDIAAIDev @LangChain

[Sam Witteveen @Sam_Witteveen]

NVIDIA released Nemotron 3 Nano Omni.

Open with a full paper of recipes for their data and SFT.

One Open model. Text, images, video, audio — all in one. Mamba-MoE backbone, Vision encoder, Parakeet audio encoder.

[Sam Witteveen @Sam_Witteveen]

7 things every team needs locked down before shipping a multi-user agent to production.

Rogue agents can run up a $10K bill overnight. Agents can hallucinate for 100s of users. Most builders don't notice. Here's what you should be monitoring.

TL;DR 📋
🔀 Model control - unified layer between your code and models, swap providers without breaking your stack
📝 Prompt registry - version your prompts, keep them out of the codebase, treat them like a second tier of code
🛡️ Guardrails - pre and post LLM, pre and post tool calls. Handle PII, PHI, prompt injection, and output filtering
��� Budget limiting - hard caps per model and project. The cloud providers won't do this for you
🔧 Tools & MCPs - central auth, granular permissions, test every one
📊 Monitoring & Tracing - trace every request, response, error, and latency spike
✅ Evals - before go-live and after. Catch the silent failures before your users do

[Sam Witteveen @Sam_Witteveen]

𝐆𝐞𝐦𝐦𝐚 𝟒 has landed with an 𝐀𝐩𝐚𝐜𝐡𝐞 𝟐.𝟎 𝐥𝐢𝐜𝐞𝐧𝐬𝐞.

That alone would be news. But what's inside these models makes it a bigger deal.
𝗧𝗟;𝗗𝗥 📝
🏗️ 4 models across 2 tiers — 𝐖𝐨𝐫𝐤𝐬𝐭𝐚𝐭𝐢𝐨𝐧 (𝟐𝟳𝗕𝗔𝟰𝗕 𝗠𝗢𝗘, 𝟯𝟭𝗕 𝗱𝗲𝗻𝘀𝗲) and 𝗘𝗱𝗴𝗲 (𝗘𝟮𝗕, 𝗘𝟰𝗕)
🧠 128 experts in the MOE, only 3.8B active per token 
🔊 Native audio support on Edge models with a 50% smaller audio encoder than Gemma 3N(681M → 305M params)
👁️ New vision encoder with variable aspect ratios and resolutions — much better OCR and document understanding
🤔 Built-in chain-of-thought 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗮𝗰𝗿𝗼𝘀𝘀 𝘁𝗲𝘅𝘁, 𝗶𝗺𝗮𝗴𝗲𝘀, 𝗮𝗻𝗱 𝗮𝘂𝗱𝗶𝗼
🛠️ 𝗙𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝗰𝗮𝗹𝗹𝗶𝗻𝗴 𝗯𝗮𝗸𝗲𝗱 𝗶𝗻 from architecture level, not prompt-coaxed
📏 𝟭𝟮𝟴𝗞 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 on edge, 𝟮𝟱𝟲𝗞 on workstation models
🌐 140 languages in pre-training, 35 in instruction tuning

[VentureBeat @VentureBeat]

Google Gemini 3.1 Pro first impressions: a 'Deep Think Mini' with adjustable reasoning on demand
https://t.co/nY9uJrPTmy

[Matt Marshall @mmarshall]

@VentureBeat @Sam_Witteveen went deep on why this acquisition is the official obituary for the ChatGPT era.

Read the breakdown on @VentureBeat here: https://t.co/yvOUOlgJab

[VentureBeat @VentureBeat]

LangChain's CEO refused to let employees install OpenClaw on company laptops — then called it one of the most important agent projects in years. The paradox explains everything about where enterprise AI is heading. https://t.co/9T36UkYGXD

[John Piazza @John_PiazzaIV]

It's true: the web is not built for AI agents. But that's changing. @Sam_Witteveen summarizes the emerging WebMCP toolset from @googlechrome, @Microsoft, et al.

Web browsing is expensive and inconsistently effective for agents. WebMCP changes that.
https://t.co/39zgvTVW0I

[Hunee @Hunee]

Big WIN for the open-source Community! Watch the full video @Eigent_AI @Sam_Witteveen

[Sam Witteveen @Sam_Witteveen]

Watch the Full video

https://t.co/rDJClLHbaY

[Sam Witteveen @Sam_Witteveen]

Anthropic launched Cowork last week.
One competing startup's response? Pivot their product and open-source EVERYTHING.

The tweet announcing it hit 1.7M views.

Here's the @Eigent_AI story 🧵

[Sam Witteveen @Sam_Witteveen]

Apple going with fine tuning of Gemini (Flash? / Flash Lite?) is probably smarter for them than jumping in to a fully "make your own models" strategy at this point in the game.

[Sam Witteveen @Sam_Witteveen]

🧵 Google just dropped UCP (Universal Commerce Protocol) at NRF with @sundarpichai himself announcing it 🛒
This is Google's play for agentic commerce - and it's built with Shopify, Etsy, Target, Walmart & Wayfair. Here's what you need to know 👇

[Sam Witteveen @Sam_Witteveen]

Happy New Year from Qwen!

[Sam Witteveen @Sam_Witteveen]

Gemini 3 Flash is here!

Google just dropped what could become your new daily driver model for AI applications. ⚡

[Sam Witteveen @Sam_Witteveen]

Anti-Vibe Coding = Pair Programming with Top Models. Good to see Junie by @jetbrains focus on this style of coding agent for their IDEs
🧵

[Google Cloud Tech @GoogleCloudTech]

@TRJ_0751 @alarcon7a @mweinbach @tabGeeks Sam Witteveen tests the new "vibe coding" capabilities in AI Studio—from cloning classic games to building slick news sites.

See what Gemini 3 built in a single shot 👇
https://t.co/C6oEJEbgKk

[Sam Witteveen @Sam_Witteveen]

Kimi K2 Thinking by @Kimi_Moonshot just dropped and it's very impressive. This model is now beating OpenAI, Anthropic, and Google on major benchmarks. The LLM landscape just shifted dramatically.

[Hunee @Hunee]

@felixhhaas Directory