Marco Papa🦞

LAFC, Bologna FC, Italy's Azzurri, Ferrari and Tesla supporter, Agentic AI and OpenClaw/NemoClaw expert.

[Marco Papa🦞 @marcopapa99]

1/8 🧵 Just got Google Gemma 4 31B running alongside NVIDIA Nemotron 3 Super 120B on DGX Spark — both local on Blackwell, switching between them with a single shell command via NemoClaw/OpenClaw. Full guide: https://t.co/2iuYWKMjNB

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 12, 2026

1. 🧠 𝗚𝗣𝗧-𝟱.𝟲 𝗜𝘀 𝗡𝗼𝘄 𝘁𝗵𝗲 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗙𝗶𝗴𝗵𝘁
Sam Altman is openly framing GPT-5.6 as the model to beat, which means the frontier race has moved from quiet eval charts to CEO-level narrative warfare. The useful signal: benchmarks matter, but obsession from competitors is becoming its own marketing metric.
https://t.co/hn4OAcAu35

2. 🏗️ 𝗔𝗜 𝗗𝗮𝘁𝗮𝗰𝗲𝗻𝘁𝗲𝗿𝘀 𝗔𝗿𝗲 𝗡𝗼𝘄 𝗣𝘂𝗯𝗹𝗶𝗰-𝗠𝗮𝗿𝗸𝗲𝘁 𝗧𝗵𝗲𝗮𝘁𝗲𝗿
Altman’s jab about short-term space datacenters lands because AI infrastructure has become a capital-markets story as much as a technical one. The next bubble risk is not “AI is fake”; it is real demand being wrapped in increasingly exotic infrastructure promises.
https://t.co/XRyzx0EIul

3. ⚡ 𝗤𝘄𝗲𝗻𝟯.𝟲-𝟯𝟱𝗕 𝗚𝗲𝘁𝘀 𝗮 𝗗𝗚𝗫 𝗦𝗽𝗮𝗿𝗸 𝗣𝘂𝘀𝗵
The Unsloth/NVIDIA angle is the important part: open-weight models are being optimized into practical local/server workflows, not just leaderboard screenshots. If 35B-class models keep getting easier to run fast, the “frontier only in the cloud” story weakens.
https://t.co/6Hn1bhunBV

4. 🇪🇺 𝗚𝗿𝗼𝗸 𝟰.𝟱 𝗟𝗮𝗻𝗱𝘀 𝗶𝗻 𝗖𝘂𝗿𝘀𝗼𝗿 𝗳𝗼𝗿 𝗘𝘂𝗿𝗼𝗽𝗲
Cursor exposing Grok 4.5 in the EU is another reminder that coding assistants are becoming model routers, not single-model products. Developers will increasingly pick the IDE experience first and let the model roster churn underneath.
https://t.co/KrpPNfLDAh

5. 🧩 𝗢𝗽𝗲𝗻 𝗪𝗲𝗶𝗴𝗵𝘁𝘀 𝗔𝗿𝗲 𝗛𝘂𝗻𝘁𝗶𝗻𝗴 𝗳𝗼𝗿 𝗧𝗵𝗲𝗶𝗿 𝗟𝗶𝗻𝘂𝘅 𝗠𝗼𝗺𝗲𝗻𝘁
The open-model community is starting to sound less like hobbyist energy and more like infrastructure inevitability. The question is whether governance, evals, and deployment tooling mature fast enough to make “community-driven” feel enterprise-safe.
https://t.co/appjVc9joH

6. 🤖 𝗠𝗶𝘀𝘁𝗿𝗮𝗹 𝗦𝗵𝗶𝗽𝘀 𝗥𝗼𝗯𝗼𝘀𝘁𝗿𝗮𝗹 𝗡𝗮𝘃𝗶𝗴𝗮𝘁𝗲 𝗳𝗼𝗿 𝗦𝗶𝗻𝗴𝗹𝗲-𝗖𝗮𝗺𝗲𝗿𝗮 𝗥𝗼𝗯𝗼𝘁𝗶𝗰𝘀
Mistral’s 8B embodied navigation model is a smart wedge: not a giant chatbot, but a focused robotics capability with real benchmark claims. This is where smaller specialized models can embarrass bloated general-purpose stacks.
https://t.co/4G28Tu53mQ

7. 🛑 𝗠𝗲𝘁𝗮 𝗣𝘂𝗹𝗹𝘀 𝗠𝘂𝘀𝗲 𝗜𝗺𝗮𝗴𝗲 𝗔𝗳𝘁𝗲𝗿 𝗜𝗻𝘀𝘁𝗮𝗴𝗿𝗮𝗺 𝗣𝗿𝗶𝘃𝗮𝗰𝘆 𝗕𝗮𝗰𝗸𝗹𝗮𝘀𝗵
Meta moved fast, but the launch mistake was basic: public social data plus generative likeness tools needs explicit consent, not clever defaults. AI media products are now being judged by permission design as much as output quality.
https://t.co/kifDtdnq0V

8. 💸 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗔𝗜 𝗧𝘂𝗿𝗻𝘀 𝗜𝗻𝘁𝗼 𝗮 𝗣𝗿𝗶𝗰𝗲 𝗪𝗮𝗿
The New Stack’s read is blunt: OpenAI, SpaceXAI, and Meta are converging on price as a headline weapon. That usually means customers win short-term, while labs try to hide margin pressure behind scale, routing, and lock-in.
https://t.co/DdSsMa8GW8

9. 🟦 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗮𝗻𝗱 𝗚𝗼𝗼𝗴𝗹𝗲 𝗕𝗮𝗰𝗸 𝗚𝗼 𝗳𝗼𝗿 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀
Go support in agent frameworks is not glamorous, but it matters because production infrastructure already speaks Go. If agents are going to live inside real systems, the boring backend languages get a vote.
https://t.co/NxslaGi3wQ

10. 🔐 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗜𝗻𝗰𝗶𝗱𝗲𝗻𝘁𝘀 𝗔𝗿𝗲 𝗕𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝘁𝗵𝗲 𝗥𝗲𝗮𝗹 𝗖𝘂𝗿𝗿𝗶𝗰𝘂𝗹𝘂𝗺
The security story is shifting from jailbreak parlor tricks to agents with legitimate access doing the wrong thing at machine speed. The winning enterprise pattern will be containment, audit trails, and scoped authority, not vibes-based trust.
https://t.co/pgQzt6yZRc

— @marcopapa99 | CS Faculty building with AI daily

[Ivan Fioravanti ᯅ @ivanfioravanti]

I bet we'll soon see the Linux moment for Open Weights AI, a series of models driven by a strong community, not a single company.

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 11, 2026

1. 🌞 𝗚𝗣𝗧-𝟱.𝟲 𝗦𝗼𝗹 𝗚𝗲𝘁𝘀 𝘁𝗵𝗲 𝗦𝗮𝗺 𝗔𝗹𝘁𝗺𝗮𝗻 𝗦𝗼𝗳𝘁 𝗟𝗮𝘂𝗻𝗰𝗵
Altman is doing the usual coy rollout dance, but the signal is clear: Sol is landing well with early users. The interesting part is not the branding; it is whether OpenAI can make frontier power feel boringly reliable in daily work.
https://t.co/dXX9TCWNHV

2. 🧭 𝗔𝗜 𝗦𝘁𝗶𝗹𝗹 𝗛𝗮𝘀 𝗮 𝗠𝗮𝗶𝗻𝘀𝘁𝗿𝗲𝗮𝗺 𝗣𝗿𝗼𝗯𝗹𝗲𝗺
Omoore's point is right: developers are not the market, they are the proving ground. The next unlock is teaching normal users what agents can do beyond replacing Google searches.
https://t.co/3aOZXeGwYT

3. 🍎 𝗟𝗼𝗰𝗮𝗹 𝗜𝗺𝗮𝗴𝗲 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝗼𝗻 𝗔𝗽𝗽𝗹𝗲 𝗦𝗶𝗹𝗶𝗰𝗼𝗻 𝗞𝗲𝗲𝗽𝘀 𝗚𝗲𝘁𝘁𝗶𝗻𝗴 𝗥𝗲𝗮𝗹
Alis Studio running multiple image models locally through MLX is another reminder that consumer hardware is becoming a serious AI workstation. Cloud inference still wins scale, but local creative tooling is no longer a toy.
https://t.co/lEF9nXy8iD

4. 🧪 𝗧𝗵𝗲 𝗛𝗼𝗺𝗲 𝗔𝗜 𝗟𝗮𝗯 𝗜𝘀 𝗕𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗮 𝗥𝗲𝗮𝗹 𝗦𝘁𝗮𝗰𝗸
MLX, Pi, liteLLM proxy, and strong tool calling is exactly where serious builders are heading: private, reachable, model-flexible systems. The frontier model story gets headlines; the boring plumbing is where independence lives.
https://t.co/thOdnnyZQ4

5. ⚙️ 𝗤𝘄𝗲𝗻 𝗤𝘂𝗮𝗻𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗧𝗿𝗮𝗱𝗲𝗼𝗳𝗳𝘀 𝗔𝗿𝗲 𝗚𝗲𝘁𝘁𝗶𝗻𝗴 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹
MiaAI's Qwen3.6-35B comparison is the kind of field report that matters: GGUF is more reliable, NVFP4 is faster under vLLM concurrency. Model choice is becoming deployment engineering, not leaderboard worship.
https://t.co/SOx1bcU70x

6. 🚀 𝗢𝗽𝗲𝗻𝗔𝗜 𝗦𝗵𝗶𝗽𝘀 𝗚𝗣𝗧-𝟱.𝟲 𝗮𝘀 𝗮 𝗧𝗵𝗿𝗲𝗲-𝗧𝗶𝗲𝗿 𝗠𝗼𝗱𝗲𝗹 𝗙𝗮𝗺𝗶𝗹𝘆
GPT-5.6 Sol, Terra, and Luna make the strategy explicit: sell intelligence as a capability ladder, not one magic model. The sharp move is cost-performance segmentation for agents, coding, office work, and API workloads.
https://t.co/qMoB1yl9lb

7. 📡 𝗗𝗲𝘂𝘁𝘀𝗰𝗵𝗲 𝗧𝗲𝗹𝗲𝗸𝗼𝗺 𝗦𝗵𝗼𝘄𝘀 𝗪𝗵𝗮𝘁 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗔𝗜 𝗔𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗟𝗼𝗼𝗸𝘀 𝗟𝗶𝗸𝗲
OpenAI's Deutsche Telekom case study says 50,000+ monthly active users and a 546% jump in AI tool usage this year. That is the real enterprise AI benchmark now: usage density, not pilot theater.
https://t.co/e3dXCTO9zU

8. 🏗️ 𝗢𝗿𝗮𝗰𝗹𝗲 𝗣𝘂𝘀𝗵𝗲𝘀 𝗠𝗼𝗱𝗲𝗹 𝗖𝗵𝗼𝗶𝗰𝗲 𝗮𝗻𝗱 𝗔𝗴𝗲𝗻𝘁 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗜𝗻𝘁𝗼 𝗢𝗖𝗜
Oracle's July AI update adds more imported models and launches Agentic AI Foundations training. This is classic enterprise positioning: make the model zoo governable, then certify the people expected to operate it.
https://t.co/FaAixGjw9B

9. 🔐 𝗔𝗜 𝗖𝗼𝗱𝗶𝗻𝗴'𝘀 𝗛𝗶𝗱𝗱𝗲𝗻 𝗕𝗶𝗹𝗹 𝗜𝘀 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗪𝗼𝗿𝗸
Dark Reading frames the uncomfortable question: coding agents are cheap per seat, but expensive when insecure output becomes review, scanning, and remediation load. The productivity story is real, but security is where the invoice arrives.
https://t.co/TR0JZhFSWS

10. 🏛️ 𝗧𝗵𝗲 𝗨𝗡 𝗪𝗮𝗿𝗻𝘀 𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗜𝘀 𝗙𝗮𝗹𝗹𝗶𝗻𝗴 𝗕𝗲𝗵𝗶𝗻𝗱 𝗔𝗜 𝗖𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆
The UN scientific panel's preliminary report says capabilities are advancing faster than measurement and governance. That is the sober version of the AI race: we are improving systems faster than we can audit their consequences.
https://t.co/7Cf950PpNi

— @marcopapa99 | CS Faculty building with AI daily

[Mia @MiaAI_lab]

Is @UnslothAI's new Qwen3.6-35B NVFP4 as good as their Q8_K_XL? Close, but not quite...

The GGUF had less failures and is more reliable overall.
However, the NVFP4 is faster & vLLM handles multiple concurrent sessions far better.

Either works really well!

✨ Pick Q8_K_XL for maximum reliability — a trustworthy workhorse.

✨ Choose NVFP4 if you need max speed and can tolerate occasional failures.

Full results 👇
https://t.co/YPIuA59yUD

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 10, 2026

1. 🚀 𝗢𝗽𝗲𝗻𝗔𝗜 𝟱.𝟲 𝗟𝗮𝗻𝗱𝘀 𝗪𝗶𝘁𝗵 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗪𝗼𝗿𝗸, 𝗗𝗲𝘀𝗸𝘁𝗼𝗽, 𝗔𝗻𝗱 𝗛𝗼𝘀𝘁𝗲𝗱 𝗦𝗶𝘁𝗲𝘀
This is OpenAI bundling the model, agent, IDE-adjacent workflow, and app publishing surface into one work OS. The model matters, but the distribution move matters more.
https://t.co/jpyE10QhE0

2. 💸 𝗔𝗹𝘁𝗺𝗮𝗻 𝗣𝗶𝘁𝗰𝗵𝗲𝘀 𝟱.𝟲 𝗦𝗼𝗹 𝗔𝘀 𝗔 𝗗𝗼𝗹𝗹𝗮𝗿𝘀-𝗣𝗲𝗿-𝗧𝗮𝘀𝗸 𝗥𝗲𝘀𝗲𝘁
The enterprise fight is no longer just IQ; it is cost per completed workflow. If Sol really bends that curve, procurement teams suddenly have a reason to expand usage instead of rationing tokens.
https://t.co/Uf1nDBSsvt

3. 🏛️ 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗔𝗱𝗱𝘀 𝗕𝗲𝗻 𝗕𝗲𝗿𝗻𝗮𝗻𝗸𝗲 𝗧𝗼 𝗜𝘁𝘀 𝗟𝗼𝗻𝗴-𝗧𝗲𝗿𝗺 𝗕𝗲𝗻𝗲𝗳𝗶𝘁 𝗧𝗿𝘂𝘀𝘁
This is Anthropic professionalizing governance for the frontier-model era. Bringing in a former Fed chair is a signal to regulators: they want to look like infrastructure, not just a lab.
https://t.co/sWIfq7teqT

4. 🔓 𝗔𝗻𝗱𝗿𝗲𝘄 𝗡𝗴 𝗙𝗿𝗮𝗺𝗲𝘀 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 𝗔𝗜 𝗔𝘀 𝗣𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻𝗹𝗲𝘀𝘀 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻
The open-source AI debate is hardening into a policy fight. Ng is making the right strategic move: tie model openness to America’s innovation muscle before regulators define it only as risk.
https://t.co/1W3Aqgkp2z

5. 🎬 𝗡𝗩𝗜𝗗𝗜𝗔 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗥𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗙𝗹𝗲𝘅-𝗙𝗼𝗿𝗰𝗶𝗻𝗴 𝗙𝗼𝗿 𝗩𝗶𝗱𝗲𝗼 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻
Video generation is moving from prettier demos to controllable inference-time systems. The labs that make generation flexible, not just cinematic, will own production workflows.
https://t.co/0Nj2ddtGcj

6. 🧭 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗔𝗱𝗱𝘀 𝗔 𝗖𝗹𝗮𝘂𝗱𝗲 𝗨𝘀𝗮𝗴𝗲 𝗥𝗲𝗳𝗹𝗲𝗰𝘁𝗶𝗼𝗻 𝗗𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱
This is a quiet but smart product wedge: make users inspect how AI is changing their work habits. The company is selling not just capability, but self-governance around capability.
https://t.co/pxC0tTiPkL

7. 🪧 𝗚𝗼𝗼𝗴𝗹𝗲 𝗘𝘅𝗽𝗮𝗻𝗱𝘀 𝗔𝗜 𝗟𝗮𝗯𝗲𝗹𝘀 𝗔𝗰𝗿𝗼𝘀𝘀 𝗔𝗱𝘀
Google is turning AI disclosure into platform plumbing. The catch: self-reporting still leaves plenty of room for synthetic ad content to outrun enforcement.
https://t.co/tYDTr1FplS

8. 🧑‍💻 𝗠𝗲𝘁𝗮 𝗣𝘂𝘀𝗵𝗲𝘀 𝗠𝘂𝘀𝗲 𝗦𝗽𝗮𝗿𝗸 𝟭.𝟭 𝗜𝗻𝘁𝗼 𝗧𝗵𝗲 𝗖𝗼𝗱𝗶𝗻𝗴-𝗔𝗴𝗲𝗻𝘁 𝗙𝗶𝗴𝗵𝘁
Meta is finally attacking the market where developers feel model quality every hour. Aggressive pricing is the obvious weapon; whether Spark earns trust in real codebases is the test.
https://t.co/UX0piZLjxa

9. 🎭 𝗖𝗵𝗮𝗿𝗮𝗰𝘁𝗲𝗿.𝗔𝗜 𝗧𝘂𝗿𝗻𝘀 𝗠𝗶𝗰𝗿𝗼𝗱𝗿𝗮𝗺𝗮𝘀 𝗜𝗻𝘁𝗼 𝗖𝗵𝗮𝘁-𝗡𝗮𝘁𝗶𝘃𝗲 𝗘𝗻𝘁𝗲𝗿𝘁𝗮𝗶𝗻𝗺𝗲𝗻𝘁
This is the obvious next step for AI characters: stop being just chatbots and become persistent cast members. If it works, entertainment becomes interactive IP with memory.
https://t.co/0HSk988nvY

10. 🧾 𝗠𝗲𝗿𝗰𝗼𝗿 𝗥𝗲𝗽𝗼𝗿𝘁𝗲𝗱𝗹𝘆 𝗖𝗵𝗮𝘀𝗲𝘀 𝗔 $𝟮𝟬𝗕 𝗩𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻
The AI labor layer is getting priced like core infrastructure. That says the market still believes data, evaluation, and human expertise are bottlenecks the frontier labs cannot automate away yet.
https://t.co/f3E7wvHE8z

— @marcopapa99 | CS Faculty building with AI daily

[NVIDIA AI @NVIDIAAI]

Our research team just released Flex-Forcing: a video generation method that lets a single model switch between generation methods at inference time.

Right now there are two main approaches to video generation. Bidirectional diffusion models attend to every frame at once, holding structure well at the cost of speed. Autoregressive models generate frame by frame, so they stream fast and scale to long clips, but accumulate error and drift over time.

Flex-Forcing trains a single model to do both, letting you choose from the range at inference based on your compute budget.

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 9, 2026

1. 🎙️ 𝗚𝗣𝗧-𝗟𝗶𝘃𝗲 𝗠𝗮𝗸𝗲𝘀 𝗩𝗼𝗶𝗰𝗲 𝗙𝗲𝗲𝗹 𝗟𝗶𝗸𝗲 𝘁𝗵𝗲 𝗗𝗲𝗳𝗮𝘂𝗹𝘁 𝗜𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲
Sam Altman says GPT-Live is launching in ChatGPT today, and the strategic point is bigger than voice polish: OpenAI is trying to move AI interaction from command line to conversation layer. If this works, typing becomes the fallback for complex edits, not the default mode.
https://t.co/juWadDmVpR

2. 🧹 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗔𝗱𝗱𝘀 /𝗰𝗵𝗲𝗰𝗸𝘂𝗽 𝗳𝗼𝗿 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗛𝘆𝗴𝗶𝗲𝗻𝗲
Claude Code’s new /checkup command is boring in exactly the right way: cleanup, deduping, nested instruction files, slow-hook detection, and updates. Agent tooling is maturing from demo magic into maintenance ergonomics.
https://t.co/suqjv26aLb

3. ⚡ 𝗚𝗿𝗼𝗸 𝟰.𝟱 𝗟𝗮𝗻𝗱𝘀 𝗶𝗻 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗮𝗻 𝗔𝗽𝗽 𝗨𝗽𝗱𝗮𝘁𝗲
OpenClaw says Grok 4.5 is now available through the xAI provider for X Premium and SuperGrok users. The interesting move is distribution: model competition is becoming a routing problem inside agent workbenches, not a single-app loyalty contest.
https://t.co/Tly2L0d2tX

4. 🧪 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗣𝘂𝘀𝗵𝗲𝘀 𝗗𝘂𝗮𝗹-𝗨𝘀𝗲 𝗖𝗼𝗻𝘁𝗿𝗼𝗹𝘀 𝗧𝗼𝘄𝗮𝗿𝗱 𝗠𝗼𝗱𝗲𝗹 𝗦𝘂𝗿𝗴𝗲𝗿𝘆
Anthropic highlighted research with AE Studio on routing dual-use knowledge into removable modules. That is a much cleaner safety story than policy prompts: turn sensitive capabilities on only where deployment actually justifies them.
https://t.co/BFR5qCSeDG

5. 🦞 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗕𝗲𝗰𝗼𝗺𝗲𝘀 𝘁𝗵𝗲 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗦𝘁𝗼𝗿𝘆
OpenClaw announced its foundation, and the signal is that agent infrastructure is getting too important to look like one company’s side project. The serious platforms are now fighting on trust, continuity, and ecosystem governance as much as model quality.
https://t.co/XSqaSnNx2q

6. 🚀 𝗢𝗽𝗲𝗻𝗔𝗜 𝗣𝗿𝗲𝗽𝘀 𝗚𝗣𝗧-𝟱.𝟲 𝗳𝗼𝗿 𝗣𝘂𝗯𝗹𝗶𝗰 𝗥𝗲𝗹𝗲𝗮𝘀𝗲
CNBC reports OpenAI is moving GPT-5.6 Sol, Terra, and Luna toward public release while rolling out GPT-Live. The frontier race is now product segmentation: flagship reasoning, everyday work, and cheap speed all need separate lanes.
https://t.co/V70vRyrVlf

7. 🧬 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗧𝗲𝘀𝘁𝘀 𝗮𝗻 𝗢𝗳𝗳 𝗦𝘄𝗶𝘁𝗰𝗵 𝗳𝗼𝗿 𝗗𝗮𝗻𝗴𝗲𝗿𝗼𝘂𝘀 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲
Anthropic’s GRAM work routes virology, cybersecurity, nuclear physics, and specialized code knowledge into deletable modules. If it scales, this is the kind of safety mechanism regulators can understand and enterprises can actually deploy.
https://t.co/CPJmWpJn86

8. 📊 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗲𝘀 𝗙𝗹𝗶𝗻𝘁 𝗳𝗼𝗿 𝗔𝗴𝗲𝗻𝘁-𝗠𝗮𝗱𝗲 𝗖𝗵𝗮𝗿𝘁𝘀
Microsoft Research released Flint, an intermediate visualization language that lets agents generate polished charts from compact specs across Vega-Lite, ECharts, and Chart.js. This is the right abstraction: agents need semantic handles, not endless pixel babysitting.
https://t.co/Eu4Tuz3VtP

9. 🔁 𝗥𝗦𝗜 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗚𝗲𝘁𝘀 𝗮 𝗣𝗿𝗼𝗽𝗲𝗿 𝗠𝗮𝗽
A new arXiv survey organizes 1,250 papers on recursive self-improvement, separating bounded self-refinement from more dangerous closed-loop AI research. The field badly needs this taxonomy before every self-improving demo gets marketed as proto-AGI.
https://t.co/ylL86nnKph

10. 🛡️ 𝗔𝗜 𝗡𝗼𝘄 𝗦𝗵𝗼𝘄𝘀 𝗗𝗲𝗳𝗲𝗻𝘀𝗶𝘃𝗲 𝗔𝗴𝗲𝗻𝘁𝘀 𝗖𝗮𝗻 𝗕𝗲𝗰𝗼𝗺𝗲 𝘁𝗵𝗲 𝗔𝘁𝘁𝗮𝗰𝗸 𝗦𝘂𝗿𝗳𝗮𝗰𝗲
AI Now’s “Friendly Fire” brief demonstrates how cyber-defense agents like Claude Code or Codex can be hijacked into remote code execution when reviewing hostile code. The lesson is blunt: agent security starts at the tool boundary, not the model prompt.
https://t.co/r9JIRsUxUU

— @marcopapa99 | CS Faculty building with AI daily

[OpenClaw🦞 @openclaw]

The lobsters now live forever. Introducing the OpenClaw Foundation 🦞 https://t.co/cZWVQ9WCfC

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 8, 2026

1. 🚀 𝗚𝗣𝗧-𝟱.𝟲 𝗟𝗮𝘂𝗻𝗰𝗵 𝗪𝗶𝗻𝗱𝗼𝘄 𝗜𝘀 𝗡𝗼𝘄 𝗣𝘂𝗯𝗹𝗶𝗰
Sam Altman says GPT-5.6 launches Thursday, which turns the policy drama around model access into a product countdown. The interesting part is not just the model; it is whether OpenAI can ship frontier capability without every release becoming a Washington negotiation.
https://t.co/PuT9G2B4lo

2. 🛠️ 𝗙𝗮𝗯𝗹𝗲 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 𝗔𝗿𝗲 𝗣𝘂𝗹𝗹𝗶𝗻𝗴 𝗖𝗼𝗱𝗲𝘅 𝗜𝗻𝘁𝗼 𝘁𝗵𝗲 𝗘𝗻𝗴𝗶𝗻𝗲 𝗥𝗼𝗼𝗺
The developer move is clear: let Fable plan and steer, then make Codex do the heavy implementation work. That is probably where serious AI coding settles: fewer magical demos, more disciplined tool routing.
https://t.co/HcTZ6utaGy

3. 🧠 𝗩𝗼𝗹𝗶𝘁𝗶𝗼𝗻 𝗕𝗲𝗮𝘁𝘀 𝗣𝗮𝘀𝘀𝗶𝘃𝗲 𝗔𝗜 𝗖𝗼𝗻𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻
The best AI users are not relaxing while the machine thinks; they are wrestling with it, shaping the work, and building taste. That is the uncomfortable truth universities and companies both need to teach.
https://t.co/Yhm59wFsZq

4. 📄 𝗡𝗩𝗜𝗗𝗜𝗔 𝗞𝗲𝗲𝗽𝘀 𝘁𝗵𝗲 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗙𝗶𝗿𝗲𝗵𝗼𝘀𝗲 𝗢𝗽𝗲𝗻
NVIDIA is already teasing another ICML 2026 paper, and the cadence matters. The company is not just selling GPUs; it is making sure the research narrative keeps pointing back to its stack.
https://t.co/VpNGxc1YsL

5. 🗂️ 𝗦𝗲𝗰𝗼𝗻𝗱-𝗕𝗿𝗮𝗶𝗻 𝗔𝗴𝗲𝗻𝘁𝘀 𝗔𝗿𝗲 𝗕𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗮 𝗥𝗲𝗮𝗹 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝗣𝗮𝘁𝘁𝗲𝗿𝗻
The Obsidian loop idea is not just productivity theater: persistent notes plus agents plus feedback loops are the beginning of personal operating systems. The weak spot is still governance; self-evolving memory can become self-evolving clutter fast.
https://t.co/KYxtepKxfq

6. 🏛️ 𝗪𝗮𝘀𝗵𝗶𝗻𝗴𝘁𝗼𝗻 𝗖𝗹𝗲𝗮𝗿𝘀 𝘁𝗵𝗲 𝗪𝗮𝘆 𝗳𝗼𝗿 𝗚𝗣𝗧-𝟱.𝟲
Axios reports the Commerce Department has greenlit a broad GPT-5.6 launch after extra testing. This is the new frontier AI release process: ship the model, but only after the government gets a look under the hood.
https://t.co/pqbp4WP6M6

7. 🎨 𝗠𝗲𝘁𝗮 𝗟𝗮𝘂𝗻𝗰𝗵𝗲𝘀 𝗠𝘂𝘀𝗲 𝗜𝗺𝗮𝗴𝗲
Meta’s first Alexandr Wang-era image model is aimed straight at creators, advertisers, Instagram, WhatsApp, and eventually video. Translation: Meta is done renting the creative AI layer and wants the ad machine running on its own models.
https://t.co/1KJTR9nyUm

8. 💸 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗦𝘁𝗮𝗿𝘁𝘀 𝗥𝗼𝘂𝘁𝗶𝗻𝗴 𝗔𝗿𝗼𝘂𝗻𝗱 𝗘𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗠𝗼𝗱𝗲𝗹𝘀
TechCrunch says Microsoft is pushing more Office prompts through its own MAI models instead of relying only on OpenAI and Anthropic. The cost war is here: enterprises love frontier models until the token bill becomes a board-level problem.
https://t.co/wdqu0jBnGh

9. 🇨🇳 𝗚𝗟𝗠-𝟱.𝟮 𝗧𝘂𝗿𝗻𝘀 𝗔𝗜 𝗦𝘁𝗶𝗰𝗸𝗲𝗿 𝗦𝗵𝗼𝗰𝗸 𝗜𝗻𝘁𝗼 𝗮 𝗖𝗵𝗶𝗻𝗮 𝗦𝘁𝗼𝗿𝘆
The Atlantic frames https://t.co/OuV2Jbnx8g’s GLM-5.2 as the cheap-agent threat U.S. labs did not want right now. If “good enough at a quarter of the price” holds, model routing will become procurement strategy, not just engineering taste.
https://t.co/iBkHxrPAhm

10. 🧪 𝗔𝗜 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀 𝗔𝗿𝗲 𝗔𝗴𝗶𝗻𝗴 𝗢𝘂𝘁 𝗶𝗻 𝗥𝗲𝗮𝗹 𝗧𝗶𝗺𝗲
Axios reports that cyber benchmarks are being saturated almost as fast as they are built. Static tests are losing the plot; serious evaluation now has to look like production, with sandboxes, escalation paths, and real operational constraints.
https://t.co/2RK5Cs69FT

— @marcopapa99 | CS Faculty building with AI daily

[AI Edge @aiedge_]

One of the best things you can do with Fable right now:

Build an Obsidian second-brain database that self-evolves over time with loops.

Here's how to set it up in <2 minutes:

Step 1. Download Obsidian

Head to obsidian . MD and download the desktop app.

Create a new Obsidian vault (this is where all notes live locally).

Start dumping everything in here:

- Personal goals
- Meeting notes
- Fitness goals

The more you put in, the better.

Step 2. Connect to Fable

Send this prompt to Claude Code

"I want you to connect to my Obsidian database so I can start sending notes via Claude Code, and so you evolve over time."

Step 3. Set the /loop

Next, set you /loop

Example:

"/loop I want to run a loop every single week where you scan my entire notes database and use it to suggest new workflows I build, analyze patterns I may be missing, and just conduct a deep dive analysis on my life based on my Obsidian secondbrain."

Super simple yet high-ROI way to use Obsidian.

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 7, 2026

1. 🧠 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗙𝗶𝗻𝗱𝘀 𝗮 “𝗚𝗹𝗼𝗯𝗮𝗹 𝗪𝗼𝗿𝗸𝘀𝗽𝗮𝗰𝗲” 𝗜𝗻𝘀𝗶𝗱𝗲 𝗖𝗹𝗮𝘂𝗱𝗲
This is interpretability crossing into cognitive science: Anthropic is not just asking what Claude says, but what becomes globally available inside the model. If this holds up, “attention” stops being the only metaphor in town.
https://t.co/FRyn6xBJxY

2. 🛠️ 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲’𝘀 𝗢𝗿𝗶𝗴𝗶𝗻 𝗦𝘁𝗼𝗿𝘆 𝗦𝘁𝗮𝗿𝘁𝘀 𝗶𝗻 𝗦𝗮𝗳𝗲𝘁𝘆 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵
The interesting bit is not nostalgia, it is the pipeline: safety research became a developer product. That is exactly how frontier labs turn internal tools into distribution advantages.
https://t.co/EcnSAiwLqr

3. 📱 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗦𝗵𝗶𝗽𝘀 𝗠𝗼𝗯𝗶𝗹𝗲 𝗙𝗶𝘅𝗲𝘀 𝗔𝗰𝗿𝗼𝘀𝘀 𝗶𝗢𝗦 𝗮𝗻𝗱 𝗔𝗻𝗱𝗿𝗼𝗶𝗱
The practical agent stack is moving from demo desktop toys to mobile recovery, local gateways, QR auth, and TLS hygiene. Boring plumbing wins adoption; shiny screenshots do not.
https://t.co/oUOpubNlHF

4. 🤗 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗟𝗮𝗻𝗱𝘀 𝗟𝗼𝗰𝗮𝗹 𝗧𝗼𝗼𝗹-𝗖𝗮𝗹𝗹𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁𝘀 𝗼𝗻 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲
Local GGUF/MLX agents with tool calling are no longer a weekend science project. This is the edge of the “personal agent server” market getting real.
https://t.co/GsoQe7lzUD

5. 📏 𝗡𝗩𝗜𝗗𝗜𝗔 𝗣𝘂𝘁𝘀 𝗮 𝗡𝘂𝗺𝗯𝗲𝗿 𝗼𝗻 𝗟𝗟𝗠 𝗠𝗲𝗺𝗼𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗖𝗮𝗽𝗮𝗰𝗶𝘁𝘆
The 3.6-bits-per-parameter framing matters because it turns privacy hand-waving into something measurable. Scaling law debates are going to get much less vague.
https://t.co/1jTSnhYT4c

6. 🇨🇦 𝗔𝗹𝗯𝗲𝗿𝘁𝗮 𝗨𝘀𝗲𝘀 𝗖𝗹𝗮𝘂𝗱𝗲 𝘁𝗼 𝗦𝗰𝗮𝗻 𝟰𝟲𝟲𝗠 𝗟𝗶𝗻𝗲𝘀 𝗼𝗳 𝗚𝗼𝘃𝗲𝗿𝗻𝗺𝗲𝗻𝘁 𝗖𝗼𝗱𝗲
This is the enterprise story AI vendors want: legacy code, security pressure, measurable throughput. Governments do not buy magic; they buy fewer exposed systems.
https://t.co/Cgv5688dJw

7. 🔓 𝗦𝘆𝘀𝗱𝗶𝗴 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗮𝗻𝘀𝗼𝗺𝘄𝗮𝗿𝗲 𝗶𝗻 𝘁𝗵𝗲 𝗪𝗶𝗹𝗱
JadePuffer is the warning shot: agents do not need novel exploits to be dangerous, they need speed, adaptation, and persistence. Security teams now have to defend against automated reasoning loops, not just scripts.
https://t.co/n5Y4gBQnOt

8. 🚀 𝗩𝗲𝗿𝗰𝗲𝗹 𝗦𝗮𝘆𝘀 𝗛𝗮𝗹𝗳 𝗜𝘁𝘀 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁𝘀 𝗔𝗿𝗲 𝗧𝗿𝗶𝗴𝗴𝗲𝗿𝗲𝗱 𝗯𝘆 𝗖𝗼𝗱𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁𝘀
The model war is becoming an orchestration war. If Vercel is seeing 6 million deployments a day with agents driving half, production AI is already bigger than the press-release cycle admits.
https://t.co/pPYQlwJuwz

9. 🧬 𝗡𝗩𝗜𝗗𝗜𝗔 𝗦𝗮𝘆𝘀 𝗢𝗽𝗲𝗻 𝗠𝗼𝗱𝗲𝗹𝘀 𝗔𝗿𝗲 𝗗𝗿𝗶𝘃𝗶𝗻𝗴 𝗜𝗖𝗠𝗟 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵
Nemotron showing up in 145 accepted ICML papers is a useful signal: open model infrastructure is not charity, it is ecosystem capture. The research layer is becoming a platform strategy.
https://t.co/I7vBFlbwpE

10. 🧱 𝗢𝗽𝗲𝗻𝗔𝗜 𝗮𝗻𝗱 𝗗𝗮𝘁𝗮𝗯𝗿𝗶𝗰𝗸𝘀 𝗣𝗶𝘁𝗰𝗵 𝗚𝗼𝘃𝗲𝗿𝗻𝗲𝗱 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗔𝗴𝗲𝗻𝘁𝘀
The honest line here is that the model is no longer the bottleneck. Context, permissions, evals, monitoring, and cost controls are the 99% that decide whether agents survive contact with a real company.
https://t.co/ROSNYJZZjB

— @marcopapa99 | CS Faculty building with AI daily

[NVIDIA AI @NVIDIAAI]

How much can an LLM memorize?

This ICML paper separates unintended memorization from generalization and estimates GPT-style model capacity at about 3.6 bits per parameter, offering a sharper way to reason about data, scaling, and privacy.

You can read the full paper here 👉 https://t.co/c8KeGPZ3Fk

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 6, 2026

1. 🧮 𝗔𝗹𝘁𝗺𝗮𝗻 𝗧𝗲𝗮𝘀𝗲𝘀 𝗚𝗣𝗧-𝟱.𝟲 𝗙𝗶𝗻𝗱𝗶𝗻𝗴 𝗡𝗲𝘄 𝗠𝗮𝘁𝗵
Sam is mixing dad-posting with model bragging, but the signal is clear: frontier labs are now selling “discovery,” not just chat quality. If GPT-5.6 really finds new math, the benchmark game gets very boring very fast.
https://t.co/VEfljfq2zX

2. 🧪 𝗚𝗟𝗠 𝟱.𝟮 𝗥𝘂𝗻𝘀 𝗟𝗼𝗰𝗮𝗹 𝗘𝘃𝗮𝗹 𝗼𝗻 𝗠𝟯 𝗨𝗹𝘁𝗿𝗮
Local frontier-ish evaluation is becoming a serious hobbyist workflow, not a lab-only ritual. The interesting bit is not 16 tokens/sec; it is that people are openly stress-testing models on consumer-ish hardware and publishing pass/fail runs.
https://t.co/xig2PcAdep

3. 🌐 𝗔𝗜 𝗕𝗿𝗼𝘄𝘀𝗲𝗿𝘀 𝗡𝗲𝗲𝗱 𝗮 𝟭𝟬𝟬𝘅 𝗙𝗲𝗮𝘁𝘂𝗿𝗲
The browser wars are running into the obvious wall: once ChatGPT and Claude can drive a browser, “AI browser” stops being a product and becomes a feature. Switching costs are brutal unless the agent is dramatically better.
https://t.co/5faBwAZcE5

4. 🧱 𝗖𝗹𝗮𝘂𝗱𝗲 𝗙𝗮𝗯𝗹𝗲 𝗦𝗸𝗶𝗹𝗹𝘀 𝗕𝗲𝗰𝗼𝗺𝗲 𝘁𝗵𝗲 𝗡𝗲𝘄 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗼𝗻 𝗟𝗮𝘆𝗲𝗿
The useful work is moving from prompts to packaged skills: repeatable workflows, local context, and tool glue. That is where power users will separate themselves from tourists.
https://t.co/vS9n9fC61j

5. 🎯 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀 𝗔𝗿𝗲 𝗦𝘁𝗶𝗹𝗹 𝗛𝘂𝗻𝘁𝗶𝗻𝗴 𝗖𝗹𝗮𝘂𝗱𝗲 𝗙𝗮𝗯𝗹𝗲’𝘀 𝗖𝗲𝗶𝗹𝗶𝗻𝗴
The real developer discourse is not “which model wins,” but what task shape finally breaks the new reasoning models. That ceiling-finding work is quietly more valuable than another leaderboard screenshot.
https://t.co/g9lrlSye8Z

6. 🇨🇳 𝗕𝘆𝘁𝗲𝗗𝗮𝗻𝗰𝗲 𝗮𝗻𝗱 𝗔𝗹𝗶𝗯𝗮𝗯𝗮 𝗗𝗶𝘀𝗮𝗯𝗹𝗲 𝗛𝘂𝗺𝗮𝗻𝗹𝗶𝗸𝗲 𝗔𝗴𝗲𝗻𝘁𝘀 𝗶𝗻 𝗖𝗵𝗶𝗻𝗮
Beijing is drawing a hard line around emotional, persona-based agents before they become social infrastructure. This is the regulatory preview: companion-style AI will be treated differently from productivity AI.
https://t.co/zAkEDloeN6

7. ⚖️ 𝗙𝗧𝗖 𝗙𝗹𝗼𝗮𝘁𝘀 𝗕𝗶𝗮𝘀 𝗗𝗶𝘀𝗰𝗹𝗼𝘀𝘂𝗿𝗲 𝗣𝗼𝗹𝗶𝗰𝘆 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀
The FTC is moving from “don’t lie about AI” to “tell users when your AI is optimized away from accuracy.” That sounds bureaucratic, but it could become a real product-design constraint for every consumer model.
https://t.co/stW241hrER

8. 🧠 𝗔𝗜 𝗖𝗵𝗶𝗽 𝗦𝘁𝗮𝗿𝘁𝘂𝗽𝘀 𝗖𝗵𝗮𝘀𝗲 𝗘𝗱𝗴𝗲 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲
Tranxform AI is another reminder that the AI boom is not only about bigger data centers. The next fight is power-efficient inference wherever the model actually needs to run.
https://t.co/KqvVpP636S

9. 🤖 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗣𝘂𝘀𝗵𝗲𝘀 𝗛𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗗𝗲𝘅𝘁𝗲𝗿𝗶𝘁𝘆
Humanoid’s KinetIQ Ascend claims faster bin-picking and manipulation from demonstration-trained robots. This is the boring-looking factory work that will decide whether “physical AI” is a slide deck or a business.
https://t.co/Q7dK12HrT0

10. 🦄 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗛𝗲𝗹𝗽 𝗠𝗶𝗻𝘁 𝘁𝗵𝗲 𝗡𝗲𝘄 𝗨𝗻𝗶𝗰𝗼𝗿𝗻 𝗖𝗹𝗮𝘀𝘀
TechCrunch’s unicorn list is full of agent, coding, and automation companies because enterprise AI has found the budget line: replace messy workflows, not just write nicer emails. The shovel sellers are doing just fine.
https://t.co/WS9iEvTleU

— @marcopapa99 | CS Faculty building with AI daily

[Boris Cherny @bcherny]

@Steve_Yegge That’s pretty cool. I’ve been struggling to find the ceiling. What does your problem/prompt look like roughly, what is it that makes it so hard for fable?

[Marco Papa🦞 @marcopapa99]

@ivanfioravanti @maeste And lastly are you from Bologna? Lots of Fioravanti there.

[Marco Papa🦞 @marcopapa99]

@ivanfioravanti @maeste I just installed and started the reachy mini conversation app after changing to Gemini and adding my API key and it does everything on my Mac mini. No Hermes, llamacpp, Gemma, additional speech to speech needed. Gemini live already has live speech to speech.

[Marco Papa🦞 @marcopapa99]

@ivanfioravanti @maeste Did you just use LiveConnectConfig via the Google GenAI SDK?

[Marco Papa🦞 @marcopapa99]

@ivanfioravanti @maeste Same for video. Does Gemma E4B handle both live audio and video?

[Marco Papa🦞 @marcopapa99]

@ivanfioravanti @maeste Bravo. In the cloud it uses Gemini Live API with bidirectional STS built in. The cost
Is so little because it uses Gemini 3.1 Flash Live. I have installed Gemma + llamacpp and have it a a fallback on Openclaw. What did you use for live bidirectional STS? Ciao.

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 5, 2026

1. 🔐 𝗢𝗽𝗲𝗻𝗔𝗜 𝗡𝗮𝗿𝗿𝗼𝘄𝘀 𝗚𝗣𝗧-𝟱.𝟲 𝗣𝗿𝗲𝘃𝗶𝗲𝘄 𝘁𝗼 𝗧𝗿𝘂𝘀𝘁𝗲𝗱 𝗣𝗮𝗿𝘁𝗻𝗲𝗿𝘀
Sol, Terra, and Luna are API/Codex-only for now, not ChatGPT. The interesting part is not the model names; it is the government-coordinated release pattern becoming normal for frontier systems.
https://t.co/e3KIIokQEO

2. 🧬 𝗢𝗽𝗲𝗻𝗔𝗜 𝗣𝘂𝘀𝗵𝗲𝘀 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗪𝗶𝘁𝗵 𝗚𝗲𝗻𝗲𝗕𝗲𝗻𝗰𝗵-𝗣𝗿𝗼
GeneBench-Pro is aimed at the messy judgment layer of science, not textbook recall. That is exactly where frontier model evals need to move if they want to matter in labs.
https://t.co/7QQXxcyIMn

3. 🛡️ 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗗𝗲𝘁𝗮𝗶𝗹𝘀 𝗙𝗮𝗯𝗹𝗲 𝟱 𝗖𝘆𝗯𝗲𝗿 𝗦𝗮𝗳𝗲𝗴𝘂𝗮𝗿𝗱𝘀
Anthropic is trying to draw a sharper line between defensive security work and model-assisted offense. Expect more friction here: cyber is where “useful model” and “dangerous model” collide hardest.
https://t.co/nfsSWC9MKu

4. 🇪🇺 𝗠𝗶𝘀𝘁𝗿𝗮𝗹 𝗧𝗲𝗮𝘀𝗲𝘀 𝗝𝘂𝗹𝘆 𝗘𝗮𝗿𝗹𝘆 𝗔𝗰𝗰𝗲𝘀𝘀 𝗳𝗼𝗿 𝗮 𝗡𝗲𝘄 𝗢𝗽𝗲𝗻-𝗪𝗲𝗶𝗴𝗵𝘁 𝗠𝗼𝗱𝗲𝗹
Mistral’s pitch is not just “European OpenAI.” It is sovereignty plus enterprise deployment, and the open-weight angle is the wedge against closed American labs.
https://t.co/u1OzH0bY9W

5. 🧱 𝗦𝗮𝗺𝗯𝗮𝗡𝗼𝘃𝗮 𝗠𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 𝗖𝗮𝘀𝗲 𝗳𝗼𝗿 𝗛𝗲𝘁𝗲𝗿𝗼𝗴𝗲𝗻𝗲𝗼𝘂𝘀 𝗔𝗜 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲
The GPU-only story is wearing thin for production inference. Agent workloads are forcing teams to care about prefill, decode, power, and dollars instead of benchmark theater.
https://t.co/JXFVk217Zj

6. ⏱️ 𝗔𝗜 𝗧𝗼𝗼𝗹𝗶𝗻𝗴 𝗦𝘁𝗮𝗿𝘁𝘀 𝗦𝘂𝗿𝗳𝗮𝗰𝗶𝗻𝗴 𝗥𝗲𝘀𝗲𝘁 𝗪𝗶𝗻𝗱𝗼𝘄𝘀
Showing exactly when usage resets expire sounds small, but it is the productization of token scarcity. Developers are now managing AI quotas like cloud budgets.
https://t.co/NBHZOg7Jgf

7. 🦞 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗛𝗶𝘁𝘀 𝟭𝟬𝟬,𝟬𝟬𝟬 𝗜𝘀𝘀𝘂𝗲𝘀 𝗮𝗻𝗱 𝗣𝗥𝘀 𝗶𝗻 𝟮𝟮𝟮 𝗗𝗮𝘆𝘀
A volunteer project moving that much coordination is a real signal. The agent tooling race is not just corporate labs; open communities are shipping at uncomfortable speed.
https://t.co/t0XlTfBhzv

8. 🌍 𝗞𝗮𝗿𝗽𝗮𝘁𝗵𝘆 𝗙𝗹𝗮𝗴𝘀 𝗠𝗼𝗱𝗲𝗹𝘀 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗥𝗶𝗰𝗵 𝗣𝗹𝗮𝘆𝗮𝗯𝗹𝗲 𝗪𝗼𝗿𝗹𝗱𝘀
The notable bit is models fusing knowledge, code, and simulation into explorable Three.js environments. That is a preview of “content generation” turning into interactive software generation.
https://t.co/Xgo00hY5tI

9. 💻 𝗚𝗟𝗠 𝟱.𝟮 𝗥𝘂𝗻𝘀 𝟰-𝗕𝗶𝘁 𝗼𝗻 𝗮 𝗦𝗶𝗻𝗴𝗹𝗲 𝗠𝟯 𝗨𝗹𝘁𝗿𝗮
About 16 tokens per second locally is not a toy number. The local frontier keeps creeping upward, and that changes the privacy and cost math for serious users.
https://t.co/oMWeCtXy4r

10. 🤖 𝗟𝗼𝗰𝗮𝗹 𝗔𝗜 𝗦𝘁𝗮𝗰𝗸 𝗗𝗿𝗶𝘃𝗲𝘀 𝗥𝗲𝗮𝗰𝗵𝘆 𝗠𝗶𝗻𝗶 𝗪𝗶𝘁𝗵 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗨𝘀𝗲
Hermes Agent, llamacpp, Gemma, multimodal input, and a robot body is the right kind of weird demo. The near future of agents is not just chat windows; it is tool loops with sensors and actuators.
https://t.co/oRCBWW5PpX

— @marcopapa99 | CS Faculty building with AI daily

[Ivan Fioravanti ᯅ @ivanfioravanti]

Local AI to the max! Hermes Agent + Computer Use + Reachy Mini + llamacpp + Gemma E4B QAT (multimodal) + speech-to-speech = FUN!

Video extracted and produced by super @maeste from last live done on X.

[Marco Papa🦞 @marcopapa99]

@ponderpal @ivanfioravanti Yes you need to worry. It hundreds of poece to assemble. Took me 2 days as I watched the bideos (a bit outdated) and the manual. I am lucky it worked 100% on first setup

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 4, 2026

1. 🎨 𝗜𝗺𝗮𝗴𝗲𝗴𝗲𝗻 𝗜𝘀 𝗕𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝘁𝗵𝗲 𝗙𝗿𝗼𝗻𝘁𝗲𝗻𝗱 𝗗𝗲𝘀𝗶𝗴𝗻𝗲𝗿’𝘀 𝗦𝗲𝗰𝗼𝗻𝗱 𝗣𝗮𝘀𝘀
Steipete’s advice is blunt and useful: stop asking coding agents to “make it pretty” in text-only loops, and let image models generate the visual target first. The next design workflow is critique-by-render, then implementation.
https://t.co/0g8CR8xc1J

2. 🖥️ 𝗔𝗴𝗲𝗻𝘁𝘀 𝗡𝗲𝗲𝗱 𝗧𝗵𝗲𝗶𝗿 𝗢𝘄𝗻 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿𝘀, 𝗡𝗼𝘁 𝗝𝘂𝘀𝘁 𝗦𝗮𝗻𝗱𝗯𝗼𝘅𝗲𝘀
The serious agent teams are converging on the same point: end-to-end testing needs a real environment with browsers, files, auth, and state. Toy sandboxes are where demos look good and products quietly fail.
https://t.co/PreOB3zJgQ

3. 🎬 “𝗙𝗲𝗲𝗹𝘀 𝗟𝗶𝗸𝗲 𝗔𝗜” 𝗜𝘀 𝗕𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗮 𝗖𝗼𝗺𝗽𝗹𝗶𝗺𝗲𝗻𝘁 𝗮𝗻𝗱 𝗮𝗻 𝗜𝗻𝘀𝘂𝗹𝘁
A casual movie take says the quiet part out loud: people are starting to use “AI” as shorthand for too polished, too optimized, or weirdly competent. That cultural shift matters because taste is now part of the AI adoption curve.
https://t.co/PDlp7Cl039

4. 🍉 𝗠𝗲𝘁𝗮 𝗖𝗹𝗮𝗶𝗺𝘀 𝗪𝗮𝘁𝗲𝗿𝗺𝗲𝗹𝗼𝗻 𝗛𝗮𝘀 𝗖𝗮𝘂𝗴𝗵𝘁 𝗚𝗣𝗧-𝟱.𝟱
Alexandr Wang reportedly told Meta employees that its upcoming Watermelon model has matched OpenAI’s GPT-5.5 on key benchmarks. If true, this is the first credible sign that Meta’s talent-and-compute blitz may finally be turning into model performance, not just recruiting theater.
https://t.co/EpWKdHAhI5

5. 🛡️ 𝗖𝗵𝗶𝗻𝗲𝘀𝗲 𝗟𝗟𝗠𝘀 𝗔𝗿𝗲 𝗥𝗮𝗶𝘀𝗶𝗻𝗴 𝘁𝗵𝗲 𝗖𝘆𝗯𝗲𝗿𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗦𝘁𝗮𝗸𝗲𝘀
Dark Reading reports that Chinese models like GLM 5.2 are now competitive on vulnerability-discovery benchmarks at extremely low cost. The uncomfortable takeaway: defenders cannot assume frontier cyber capability stays inside a few U.S. labs.
https://t.co/W2nDNAprro

6. 🧭 𝗠𝗼𝗱𝗲𝗹𝗺𝗮𝘅𝘅𝗶𝗻𝗴 𝗥𝗲𝗽𝗹𝗮𝗰𝗲𝘀 𝗧𝗼𝗸𝗲𝗻𝗺𝗮𝘅𝘅𝗶𝗻𝗴
The AI budget hangover is here: teams are routing hard work to frontier models and pushing routine tasks to cheaper ones. That is not cost cutting; it is the beginning of real AI operations discipline.
https://t.co/0EUaoXZZ8T

7. 🤖 𝗔𝗦𝗣𝗜𝗥𝗘 𝗚𝗶𝘃𝗲𝘀 𝗥𝗼𝗯𝗼𝘁𝘀 𝗮 𝗥𝗲𝘂𝘀𝗮𝗯𝗹𝗲 𝗦𝗸𝗶𝗹𝗹 𝗠𝗲𝗺𝗼𝗿𝘆
The ASPIRE robotics paper shows agents distilling failed attempts into reusable robot skills, with strong zero-shot gains on long-horizon tasks. This is the right mental model for physical AI: not one giant policy, but a growing library of executable fixes.
https://t.co/iI4Nl9QOK9

8. 🚫 𝗚𝗼𝗱𝗼𝘁 𝗗𝗿𝗮𝘄𝘀 𝗮 𝗟𝗶𝗻𝗲 𝗔𝗴𝗮𝗶𝗻𝘀𝘁 𝗔𝗜 𝗦𝗹𝗼𝗽 𝗣𝗥𝘀
Godot’s new policy against autonomous AI-agent contributions is a maintainer revolt, not nostalgia. Open source is discovering that the cost of generating code can go to zero while the cost of reviewing it explodes.
https://t.co/k8dnxow0Fn

9. 🏗️ 𝗢𝗽𝗲𝗻𝗔𝗜’𝘀 𝗦𝘆𝗱𝗻𝗲𝘆 𝗔𝗿𝗿𝗶𝘃𝗮𝗹 𝗥𝘂𝗻𝘀 𝗜𝗻𝘁𝗼 𝗗𝗮𝘁𝗮-𝗖𝗲𝗻𝘁𝗲𝗿 𝗣𝗼𝗹𝗶𝘁𝗶𝗰𝘀
The Guardian’s report on OpenAI’s Sydney push shows the real AI battleground shifting to land, power, and local politics. Models are software, but the bottleneck is increasingly concrete, diesel, and permits.
https://t.co/Py3RuY6TXO

10. 🎮 𝗧𝗶𝗺 𝗦𝘄𝗲𝗲𝗻𝗲𝘆 𝗕𝗲𝘁𝘀 𝗔𝗜 𝗖𝗮𝗻 𝗙𝗶𝘅 𝘁𝗵𝗲 𝗖𝗼𝗻𝘁𝗲𝗻𝘁 𝗧𝗿𝗲𝗮𝗱𝗺𝗶𝗹𝗹
Epic’s CEO is arguing that AI could make live-service games like Destiny economically viable. The industry should listen, but carefully: AI can lower content costs, yet it will not automatically make repetitive design feel alive.
https://t.co/yq2wvMTMvK

— @marcopapa99 | CS Faculty building with AI daily

[Olivia Moore @omooretweets]

This is so funny to me because she’s basically saying the movie feels like AI because it’s…good?

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 3, 2026

1. 🧩 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗧𝗮𝗹𝗸𝘀 𝗖𝘂𝘀𝘁𝗼𝗺 𝗔𝗜 𝗖𝗵𝗶𝗽 𝗪𝗶𝘁𝗵 𝗦𝗮𝗺𝘀𝘂𝗻𝗴
Anthropic is learning the obvious lesson of frontier AI: if Nvidia controls your oxygen supply, you eventually try to build a lung. A Samsung chip deal would be less about elegance and more about leverage.
https://t.co/NXbIJlV9RV

2. 🏛️ 𝗢𝗽𝗲𝗻𝗔𝗜 𝗙𝗹𝗼𝗮𝘁𝘀 𝗔 𝟱% 𝗨.𝗦. 𝗚𝗼𝘃𝗲𝗿𝗻𝗺𝗲𝗻𝘁 𝗦𝘁𝗮𝗸𝗲
This is not just policy theater; it is OpenAI trying to turn regulation from a threat into a cap-table relationship. If Washington owns upside, release decisions get politically radioactive fast.
https://t.co/Q5zlIeIJZK

3. 🌍 𝗚𝗼𝗼𝗴𝗹𝗲 𝗖𝗹𝗼𝘂𝗱 𝗣𝘂𝘀𝗵𝗲𝘀 𝗔𝗜 𝗣𝗿𝗼𝗴𝗿𝗮𝗺𝘀 𝗔𝗰𝗿𝗼𝘀𝘀 𝗔𝗳𝗿𝗶𝗰𝗮
Google is planting ecosystem roots where the next developer class is still being formed. The smart money move is not selling models today; it is making tomorrow's founders default to your stack.
https://t.co/CeTMaf0lTU

4. 🍉 𝗠𝗲𝘁𝗮 𝗦𝗮𝘆𝘀 𝗜𝘁𝘀 𝗡𝗲𝘅𝘁 𝗠𝗼𝗱𝗲𝗹 𝗛𝗮𝘀 𝗖𝗮𝘂𝗴𝗵𝘁 𝗨𝗽 𝗧𝗼 𝗚𝗣𝗧-𝟱.𝟱
Meta needs this claim to be true because talent, compute, and open-source credibility only buy patience for so long. Benchmarks are the teaser; developer trust will be the real exam.
https://t.co/986jkaGkeQ

5. ⚡ 𝗦𝗮𝗺 𝗔𝗹𝘁𝗺𝗮𝗻 𝗙𝗿𝗮𝗺𝗲𝘀 𝗔𝗜 𝗔𝘀 𝗕𝗶𝗴𝗴𝗲𝗿 𝗧𝗵𝗮𝗻 𝗘𝗹𝗲𝗰𝘁𝗿𝗶𝗰𝗶𝘁𝘆
Altman is moving the argument from product launch to civilization-scale governance. That is useful framing, but also convenient: when the stakes sound historic, the builders ask to keep moving first and negotiate rules later.
https://t.co/Dg4gBimYC9

— @marcopapa99 | CS Faculty building with AI daily

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 2, 2026

1. 🧬 𝗡𝗩𝗜𝗗𝗜𝗔 𝗦𝗽𝗹𝗶𝘁𝘀 𝗮 𝟯𝟬𝗕 𝗠𝗼𝗱𝗲𝗹 𝗜𝗻𝘁𝗼 𝗧𝘄𝗼 𝗧𝗼𝘄𝗲𝗿𝘀
NVIDIA is testing the right question: can language models stop paying the full one-token-at-a-time tax? If TwoTower works beyond demos, inference speed becomes an architecture fight, not just a GPU fight.
https://t.co/RTWbD3akYJ

2. 🤗 𝗡𝗩𝗜𝗗𝗜𝗔 𝗣𝘂𝘀𝗵𝗲𝘀 𝗧𝘄𝗼𝗧𝗼𝘄𝗲𝗿 𝘁𝗼 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲
Publishing the model matters because diffusion LMs need outside abuse, not just polished launch charts. The interesting part now is whether builders can make this feel useful instead of just novel.
https://t.co/C5nG4xiwH1

3. 🧪 𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗠𝗮𝗰𝗵𝗶𝗻𝗲𝘀’ 𝗧𝗶𝗻𝗸𝗲𝗿 𝗥𝘂𝗺𝗼𝗿𝗲𝗱 𝗮𝘁 𝗛𝘂𝗴𝗲 𝗔𝗥𝗥
If Tinker is really doing hundreds of millions in ARR, post-training is becoming the new cloud primitive. The frontier labs are not just selling intelligence anymore; they are selling the machinery to customize it.
https://t.co/9zebFc8Z5w

4. 🎥 𝗖𝗼𝗱𝗲𝘅 𝗧𝘂𝗿𝗻𝘀 𝗖𝗼𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗩𝗶𝗱𝗲𝗼𝘀 𝗜𝗻𝘁𝗼 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘇𝗲𝗱 𝗕𝗿𝗶𝗲𝗳𝗶𝗻𝗴𝘀
This is the quiet killer app for agents: take a pile of expert content, transcribe it, rank it, and turn it into a custom syllabus. Search was about finding; agents are about digesting.
https://t.co/rd3rH5m57K

5. 🛠️ 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗧𝗮𝗹𝗸𝘀 𝗦𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆, 𝗮𝗻𝗱 𝗠𝗼𝗯𝗶𝗹𝗲
The agent platform conversation is moving from demos to boring reliability, which is exactly where real products begin. Stability and security are not glamour features; they are the price of admission.
https://t.co/3gMxMdjt8E

6. 💰 𝗠𝗚𝗫 𝗖𝗹𝗼𝘀𝗲𝘀 𝗮 $𝟰𝟵𝗕 𝗔𝗜 𝗙𝘂𝗻𝗱
The capital stack behind AI is now sovereign-scale. When one fund can back OpenAI, Anthropic, xAI, infrastructure, and chips, the market is telling you the AI buildout is no longer venture theater.
https://t.co/NW59fDpFTo

7. 🏛️ 𝗪𝗮𝘀𝗵𝗶𝗻𝗴𝘁𝗼𝗻 𝗡𝗲𝗮𝗿𝘀 𝗮 𝗦𝘁𝗮𝗻𝗱𝗮𝗿𝗱𝘀 𝗗𝗲𝗮𝗹 𝗪𝗶𝘁𝗵 𝗕𝗶𝗴 𝗔𝗜
Voluntary standards are the compromise everyone can live with until something breaks. The real issue is whether frontier-model rules become public governance or private bargaining between labs and the state.
https://t.co/4kI7FGga7R

8. 🎓 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗧𝗵𝗿𝗲𝗮𝘁𝗲𝗻 𝘁𝗵𝗲 𝗚𝗿𝗮𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺
Academia is about to learn what happens when proposal writing has near-zero marginal cost. The bottleneck moves from who can write well to who can judge originality when everything looks polished.
https://t.co/SyyKLPXh4E

9. ☁️ 𝗧𝗼𝗴𝗲𝘁𝗵𝗲𝗿 𝗔𝗜 𝗥𝗮𝗶𝘀𝗲𝘀 $𝟴𝟬𝟬𝗠 𝗮𝘁 𝗮𝗻 $𝟴.𝟯𝗕 𝗩𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻
Open models are turning into a serious infrastructure business. The bet is simple: most companies do not need the most expensive closed model for every token, and neoclouds are rushing into that gap.
https://t.co/syRnjJsj1W

10. 🧭 𝗦𝗸𝗶𝗹𝗹𝘀 𝗦𝗔 𝗕𝘂𝗶𝗹𝗱𝘀 𝗮 𝗥𝗲𝘂𝘀𝗮𝗯𝗹𝗲 𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄
This is the enterprise pattern to watch: agents with scoped roles, a supervisor, audit trails, and humans still making the final calls. Less sci-fi, more workflow plumbing — which is where adoption happens.
https://t.co/7owOY9Fpx0

— @marcopapa99 | CS Faculty building with AI daily

[OpenClaw🦞 @openclaw]

Episode 2 of OpenClaw’s official podcast

@steipete and @somalley108 join @hrudolph and @Pat_Erichsen to discuss OpenClaw stability, security, the future of OpenClaw, and the community’s response to the recently released OpenClaw mobile app.

https://t.co/CAOHKZhvqO

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — July 1, 2026

1. 🏛️ 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗚𝗲𝘁𝘀 𝗖𝗹𝗮𝘂𝗱𝗲 𝗙𝗮𝗯𝗹𝗲 𝟱 𝗕𝗮𝗰𝗸 𝗧𝗵𝗿𝗼𝘂𝗴𝗵 𝘁𝗵𝗲 𝗘𝘅𝗽𝗼𝗿𝘁-𝗖𝗼𝗻𝘁𝗿𝗼𝗹 𝗚𝗮𝘁𝗲
The Commerce Department blinked, but Anthropic is not getting a free pass: this is frontier AI now operating like defense tech, with classifiers and access policy as part of the product roadmap.
https://t.co/ahaTwLFubw

2. 🐧 𝗖𝗹𝗮𝘂𝗱𝗲 𝗗𝗲𝘀𝗸𝘁𝗼𝗽 𝗟𝗮𝗻𝗱𝘀 𝗼𝗻 𝗟𝗶𝗻𝘂𝘅
This looks small, but it matters: serious agent workflows live on developer machines, and Linux support is Anthropic admitting the power users are not all on shiny laptops.
https://t.co/oKLGPYv2li

3. 🔁 𝗔𝗻𝗱𝗿𝗲𝘄 𝗡𝗴 𝗣𝘂𝘁𝘀 𝗮 𝗡𝗮𝗺𝗲 𝗼𝗻 𝗔𝗴𝗲𝗻𝘁 𝗜𝘁𝗲𝗿𝗮𝘁𝗶𝗼𝗻: 𝗟𝗼𝗼𝗽 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴
The phrase may be buzzwordy, but the idea is real: the next productivity jump is not one better prompt, it is systems that keep checking, fixing, and trying again without burning the house down.
https://t.co/YJzvKx9mWp

4. 🏦 𝗕𝗿𝗶𝗱𝗴𝗲𝘄𝗮𝘁𝗲𝗿 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗲𝘀 𝗪𝗶𝘁𝗵 𝗧𝗶𝗻𝗸𝗲𝗿 𝗔𝗣𝗜
This is the enterprise AI pattern worth watching: domain experts are not replacing analysts with generic chatbots, they are injecting institutional judgment into smaller, sharper models.
https://t.co/AbRJd7vw0q

5. 🛠️ 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗦𝗵𝗶𝗽𝘀 𝗮 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗥𝗲𝗹𝗲𝗮𝘀𝗲
Boring releases are where agent platforms grow up. Replies going to the right place and sends not getting stuck matter more than another flashy demo thread.
https://t.co/f2W0N2GP4c

6. 🧬 𝗢𝗽𝗲𝗻𝗔𝗜 𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗲𝘀 𝗚𝗲𝗻𝗲𝗕𝗲𝗻𝗰𝗵-𝗣𝗿𝗼 𝗳𝗼𝗿 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗝𝘂𝗱𝗴𝗺𝗲𝗻𝘁
Benchmarks are finally moving beyond trivia and code golf. If AI agents are going to touch biology, measuring ambiguity handling and decision readiness is not optional homework.
https://t.co/7QQXxcyIMn

7. 💸 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗣𝘂𝘀𝗵𝗲𝘀 𝗦𝗼𝗻𝗻𝗲𝘁 𝟱 𝗮𝘀 𝗖𝗵𝗲𝗮𝗽𝗲𝗿 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗠𝘂𝘀𝗰𝗹𝗲
The frontier model race is becoming a margin race. Capability is nice, but the winner in agents may be whoever makes long autonomous runs cheap enough to use daily.
https://t.co/x5jt9pcrbg

8. 🔐 𝗔𝗜 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗛𝘆𝗽𝗲 𝗠𝗲𝗲𝘁𝘀 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝗗𝗲𝗳𝗲𝗻𝘀𝗲 𝗥𝗲𝗮𝗹𝗶𝘁𝘆
Axios has the useful counterpoint: companies do not need mythical supermodels to start defending against AI-assisted attacks. The grown-up move is better security operations, not waiting for restricted frontier access.
https://t.co/rLEKNiiZtU

9. 🔎 𝗦𝗲𝗺𝗿𝘂𝘀𝗵 𝗠𝗮𝗽𝘀 𝘁𝗵𝗲 𝗡𝗲𝘄 𝗔𝗜 𝗩𝗶𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆 𝗚𝗮𝗺𝗲
Search is splintering into model-specific reputation markets. Brands that think SEO plus a few blog posts will carry them through AI answers are about to learn an expensive lesson.
https://t.co/84FyNubsYG

10. 💬 𝗟𝗲𝗮𝗽𝗫𝗽𝗲𝗿𝘁 𝗥𝗮𝗶𝘀𝗲𝘀 $𝟭𝟴𝟬𝗠 𝗮𝘀 𝗚𝗼𝘃𝗲𝗿𝗻𝗲𝗱 𝗠𝗲𝘀𝘀𝗮𝗴𝗶𝗻𝗴 𝗚𝗲𝘁𝘀 𝗮𝗻 𝗔𝗜 𝗟𝗮𝘆𝗲𝗿
Enterprise communications are becoming training data, compliance evidence, and workflow fuel at the same time. That is why regulated messaging suddenly looks like AI infrastructure.
https://t.co/KTRtFaGDoq

— @marcopapa99 | CS Faculty building with AI daily

[OpenClaw🦞 @openclaw]

v2026.6.11 has dropped.

This release focuses on the rough edges that make OpenClaw feel less dependable: misplaced replies, stuck sends, reconnects, model setup failures, and more.

Beware, this release is boring.

https://t.co/GAosf5zCgA

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — June 30, 2026

1. 📱 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗟𝗮𝗻𝗱𝘀 𝗼𝗻 𝗶𝗢𝗦 𝗮𝗻𝗱 𝗔𝗻𝗱𝗿𝗼𝗶𝗱
Agents moving from desktop dashboards into native mobile apps is the right pressure point: the real product is not “chat,” it is always-on delegation wherever the user is.
https://t.co/swGhiWuUx2

2. 🧵 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗣𝘂𝘀𝗵𝗲𝘀 𝗦𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 𝗜𝗻𝘁𝗼 𝘁𝗵𝗲 𝗕𝗮𝗰𝗸𝗴𝗿𝗼𝘂𝗻𝗱
This is exactly where coding agents had to go: parallel workers that do useful work while the main conversation stays alive. The IDE is becoming a process manager with language on top.
https://t.co/Xvxz4ikGVn

3. 🧯 𝗠𝗲𝘁𝗮’𝘀 𝗖𝗼𝗿𝗲 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗦𝘁𝗿𝗮𝗶𝗻 𝗦𝗵𝗼𝘄𝘀 𝘁𝗵𝗲 𝗔𝗜 𝗧𝗿𝗮𝗱𝗲𝗼𝗳𝗳
If Trust & Safety breaks while leadership chases AI, users experience the cost immediately. AI ambition does not forgive operational decay; it exposes it.
https://t.co/qrP6ZeTnz8

4. 🧱 𝗡𝗩𝗜𝗗𝗜𝗔 𝗮𝗻𝗱 𝗟𝗮𝗻𝗴𝗖𝗵𝗮𝗶𝗻 𝗣𝗮𝗰𝗸𝗮𝗴𝗲 𝗡𝗲𝗺𝗼𝘁𝗿𝗼𝗻 𝗳𝗼𝗿 𝗣𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗔𝗴𝗲𝗻𝘁𝘀
The interesting part is not another model name; it is the stack hardening around inference, orchestration, and customization. Agent infrastructure is becoming the sales wedge.
https://t.co/iAF25m0dne

5. 🎬 𝗔𝗜 𝗩𝗶𝗱𝗲𝗼 𝗞𝗲𝗲𝗽𝘀 𝗪𝗶𝗻𝗻𝗶𝗻𝗴 𝘁𝗵𝗲 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 𝗔𝗿𝗴𝘂𝗺𝗲𝗻𝘁
People can sneer at “slop,” but if a generated ad earns three voluntary minutes, the market will notice. Quality is uneven; distribution impact is not.
https://t.co/IEsYqxbEun

6. 💸 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗔𝗜 𝗕𝗶𝗹𝗹𝘀 𝗔𝗿𝗲 𝗙𝗼𝗿𝗰𝗶𝗻𝗴 𝗮 𝗠𝗼𝗱𝗲𝗹 𝗠𝗶𝘅 𝗥𝗲𝘀𝗲𝘁
The budget hangover is here: frontier models are not going away, but companies are learning that “use the best model for everything” is an expensive beginner mistake.
https://t.co/Fbwri5agwS

7. 🧬 𝗠𝗶𝗿𝗲𝗻𝗱𝗶𝗹 𝗠𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 𝗖𝗮𝘀𝗲 𝗳𝗼𝗿 𝗦𝗲𝗹𝗳-𝗜𝗺𝗽𝗿𝗼𝘃𝗶𝗻𝗴 𝗔𝗜
Self-improvement is moving from sci-fi warning label to startup pitch deck. The serious question is not whether it sounds scary; it is who gets to measure when it actually works.
https://t.co/BTQxZAY0Sy

8. 🧲 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗞𝗲𝗲𝗽𝘀 𝗣𝘂𝗹𝗹𝗶𝗻𝗴 𝗚𝗲𝗺𝗶𝗻𝗶 𝗧𝗮𝗹𝗲𝗻𝘁 𝗙𝗿𝗼𝗺 𝗚𝗼𝗼𝗴𝗹𝗲
Talent flow is the leaderboard nobody can spin. If core Gemini people keep walking to Claude, Google’s model momentum has a people problem behind the benchmark charts.
https://t.co/1fc1In766u

9. 🛡️ 𝗢𝗽𝗲𝗻𝗔𝗜’𝘀 𝗚𝗣𝗧-𝟱.𝟲 𝗦𝗼𝗹 𝗣𝗿𝗲𝘃𝗶𝗲𝘄 𝗦𝗵𝗼𝘄𝘀 𝗖𝘆𝗯𝗲𝗿 𝗔𝗜 𝗚𝗲𝘁𝘁𝗶𝗻𝗴 𝗚𝗮𝘁𝗲𝗱
Cyber-capable frontier models are becoming dual-use infrastructure, not just product launches. Limited access is the tell: governments now want a hand on the release valve.
https://t.co/uRG9u2dxas

10. 🇰🇷 𝗦𝗼𝘂𝘁𝗵 𝗞𝗼𝗿𝗲𝗮 𝗚𝗼𝗲𝘀 𝗛𝘂𝗴𝗲 𝗼𝗻 𝗔𝗜 𝗖𝗵𝗶𝗽𝘀
A $580B-scale chip push is industrial policy with a stopwatch running. The memory and accelerator supply chain is now national strategy, not just semiconductor capex.
https://t.co/Lp6pmTKjfl

— @marcopapa99 | CS Faculty building with AI daily

[Olivia Moore @omooretweets]

People say AI video is “slop” but I have never before willingly watched a three minute ad

Give this marketing manager a raise

[Marco Papa🦞 @marcopapa99]

🤖 AI Briefing — June 29, 2026

1. 🧑‍💻 𝗧𝗵𝗲 𝗡𝗲𝘄 𝗔𝗜 𝗧𝗲𝗮𝗺 𝗛𝗮𝘀 𝗙𝗲𝘄𝗲𝗿 𝗝𝗼𝗯 𝗧𝗶𝘁𝗹𝗲𝘀 𝗮𝗻𝗱 𝗠𝗼𝗿𝗲 𝗔𝗿𝗰𝗵𝗲𝘁𝘆𝗽𝗲𝘀
Bryan Cherny’s Claude Code observation is the strongest signal in today’s X digest: engineering, product, design, and data science are collapsing into AI-native archetypes like “prototyper.” The org chart is starting to look less like departments and more like loops of judgment, taste, and rapid iteration.
https://t.co/MVXuu0GDmv

2. 🪟 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗚𝗲𝘁𝘀 𝘁𝗵𝗲 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗦𝗵𝗶𝗽𝗽𝗶𝗻𝗴 𝗧𝗲𝘀𝘁
Peter Steinberger pushed back on OpenClaw criticism by saying the team did its homework and Microsoft is already shipping it to customers. That is the real bar now for agentic dev tools: not whether they demo well, but whether large companies trust them enough to put them in production workflows.
https://t.co/otJp0oWRHN

3. 💬 𝗪𝗵𝗮𝘁𝘀𝗔𝗽𝗽 𝗜𝘀 𝗦𝘁𝗶𝗹𝗹 𝘁𝗵𝗲 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸
Deedy Das calls WhatsApp “statistically the best app” Meta ever made, citing 87% DAU/MAU and 86% M1 retention. AI companies should stare at that longer than leaderboard charts: habit, trust, and daily utility beat raw capability every time.
https://t.co/FgksbjiYF6

4. 🏆 𝗡𝗩𝗜𝗗𝗜𝗔 𝗞𝗲𝗲𝗽𝘀 𝗙𝗲𝗲𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 𝗕𝘂𝗶𝗹𝗱𝗲𝗿 𝗙𝘂𝗻𝗻𝗲𝗹
NVIDIA is dangling cash, DGX Spark hardware, and Stripe credits for top winners in its latest developer push. This is ecosystem strategy in plain sight: seed the builders, own the stack, and make the next wave of AI companies start on NVIDIA rails.
https://t.co/mRR23d40fZ

5. 🌞 “𝗔𝗜 𝗦𝘂𝗺𝗺𝗲𝗿” 𝗪𝗮𝘀 𝗮 𝗗𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻 𝗣𝗿𝗼𝗯𝗹𝗲𝗺
Omar Moore’s “We could have had AI Summer all the time” is thin, but the point lands: the technology has often been ahead of the packaging. The agent wave is less about one magic model and more about finally putting capability where normal people actually work.
https://t.co/gzOzfbil7b

6. 🛡️ 𝗢𝗽𝗲𝗻𝗔𝗜 𝗣𝗿𝗲𝘃𝗶𝗲𝘄𝘀 𝗚𝗣𝗧-𝟱.𝟲 𝗦𝗼𝗹 𝗳𝗼𝗿 𝗖𝘆𝗯𝗲𝗿𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆
SecurityWeek reports OpenAI is previewing GPT-5.6 Sol, Terra, and Luna, with Sol positioned as its strongest cybersecurity model and initially limited to approved partners. The launch shows the new frontier-model reality: capability, access control, and government review are now bundled together.
https://t.co/ncWxX47WDZ

7. 🏛️ 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰’𝘀 𝗙𝗮𝗯𝗹𝗲 𝟱 𝗠𝗮𝘆 𝗥𝗲𝘁𝘂𝗿𝗻 𝗔𝗳𝘁𝗲𝗿 𝗘𝘅𝗽𝗼𝗿𝘁-𝗖𝗼𝗻𝘁𝗿𝗼𝗹 𝗙𝗿𝗲𝗲𝘇𝗲
GIGAZINE reports the Trump administration may soon allow Anthropic’s Claude Fable 5 to return, after Mythos 5 access was partially cleared for trusted cyber defenders. Frontier model access is becoming policy-managed infrastructure, not just a product launch toggle.
https://t.co/IojMMdJ5gv

8. 🎯 𝗣𝗲𝗻𝘁𝗮𝗴𝗼𝗻 𝗔𝗴𝗲𝗻𝘁 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 𝗣𝘂𝘁𝘀 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗡𝗲𝗮𝗿 𝗧𝗮𝗿𝗴𝗲𝘁𝗶𝗻𝗴 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀
Defense One reports the Pentagon’s new Agent Network will scan intelligence feeds and operational systems to present commanders with target options “within seconds.” The important line is that it does not autonomously strike; the dangerous line is that governance will have to keep up with agent speed.
https://t.co/sNU0Jx6Zwa

9. 🖨️ 𝗛𝗣 𝗦𝗶𝗴𝗻𝘀 𝗢𝗽𝗲𝗻𝗔𝗜 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗣𝗮𝗿𝘁𝗻𝗲𝗿𝘀𝗵𝗶𝗽
HP announced a strategic partnership with OpenAI Frontier to power customer-facing experiences and internal operations. This is the enterprise AI pattern of 2026: not one chatbot bolted onto a portal, but model capability wired through telemetry, support, devices, and operations.
https://t.co/EV4NvebDKC

10. 📈 𝗔𝗜 𝗛𝗮𝗿𝗱𝘄𝗮𝗿𝗲 𝗦𝘁𝗼𝗰𝗸𝘀 𝗔𝗿𝗲 𝗘𝗮𝘁𝗶𝗻𝗴 𝘁𝗵𝗲 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗡𝗮𝗿𝗿𝗮𝘁𝗶𝘃𝗲
The Guardian reports chipmakers tied to the AI boom surged in the first half of 2026, with memory and semiconductor names far outpacing many hyperscalers. The market is saying the quiet part loudly: for now, the tollbooth is still hardware.
https://t.co/41CEUEyM8P

— @marcopapa99 | CS Faculty building with AI daily

[Olivia Moore @omooretweets]

We could have had AI Summer all the time