Billy Lee

The most underrated AI announcement this week wasn’t a model. It was Microsoft building a sandbox for AI agents. That sounds boring until you realize: agents are moving from “answer my question” to “go do the task.” And once AI can act, security becomes the product. At Build 2026, Microsoft announced Microsoft Execution Containers, or MXC. The idea: give agents controlled environments where their actions can be contained, governed, and attributed. Not just “trust the model.” More like: “give the agent a badge, a room, rules, and an audit trail.” That matters because the next wave of AI is not just smarter text. It’s agents that can write code, change files, call tools, and operate across apps. A normal user should care because this is the difference between: “AI suggested something” and “AI changed something on my behalf.” This is where the agent race gets real. The winners won’t just have the flashiest model. They’ll have the boring infrastructure that makes people comfortable letting AI touch actual work. Permissions. Identity. Sandboxing. Logs. That’s the trust layer. Sources: Microsoft official blog: blogs.microsoft.com/blog/2026/06/0… Windows Developer Blog: blogs.windows.com/windowsdevelop… Microsoft Foundry Blog: devblogs.microsoft.com/foundry/build-…

@DustyBC Nice to see it actually advancing. Clarity on this stuff helps everyone building

🚨𝗝𝗨𝗦𝗧 𝗜𝗡: 🇺🇸 Senator Cynthia Lummis says “The Clarity Act passed committee. The floor is next. We did not come this far to quit at the 5 yard line.”

@zephyr_z9 85k just for the memory is wild. Local AI is fun until you see the actual bill

$85k for 250 Gigs of HBM3E Seeing stuff like this always makes me super bearish on edge AI/self-hosting

@chekanboy @RialoHQ @keepyourpixel @LongLifeFox @ardortester Exactly, sounds like giving a stranger your passwords and hoping for the best

Imagine giving a new guy access to your bank account > your company card > your clients > your passwords and telling him: "just do whatever you think is right" Sounds insane yet that’s basically how ppl deploy AI agents today Crypto spent years teaching us: don’t give strangers your private keys Now people are giving AI agents access to wallets and calling it innovation lmao Feels like we’re about to relearn the same lesson with better marketing @RialoHQ will fix it

@FT Agree, personhood for agents opens up a whole can of worms legally

We must not grant AI agents legal personhood ft.trib.al/BnEez25

@WallStreetApes Yeah the noise from those turbines sounds brutal. Wonder if they'll end up buying people out or just keep expanding anyway

Homeowners say they’re having a hard time living next to the new massive OpenAI Stargate AI data center in Abilene, Texas They say the noise is unbearable due to the 10 turbines used for power The Stargate data center is planning on installing another 41 additional turbines which will greatly add to the noise The homeowners feel trapped and are brought to tears. This woman and her husband are veterans The site will 51 gas turbines plus dozens of diesel generators for backup I looked it up and found the amount of noise this would produce once done is equivalent to a food processor at all times blasting into people’s homes

Ran into this wild thread about a tiny Japanese company called Tri Chemical Laboratories that basically owns the secret sauce for making the ultra-pure chemicals needed in every cutting-edge AI chip. We're talking atomic layer stuff for sub-3nm processes, and they have patents that make switching suppliers a nightmare for the big fabs. Samsung even tried and lost in court. The margins are insane too. Feels like one of those quiet picks-and-shovels plays nobody talks about until the supply chain hiccups. x.com/i/status/20635… Then there's this solo founder who basically turned Claude into an entire company's operating system in seven days. Every department, every SOP, every permission set, all connected in this living graph where you just click around and the right context pops up. Way past Notion or whatever. Makes you wonder how many teams are about to get replaced by one person with good prompts and a terminal. x.com/i/status/20635… Anthropic quietly filed for an IPO while people are still arguing about safety limits on self-improving models. Google dropped Gemma 4 12B and OpenAI is adding more hosted tools. Feels like the pace is only speeding up, especially with all the agent stuff and physical world models floating around. The gap between what one motivated person can ship now and what used to take a whole funded team is getting ridiculous. x.com/Agos_Labs/stat… What's the most surprising solo project you've seen lately, or are you more interested in the hardware side of all this AI growth? Follow for more daily drops like this.

@satyaxtwt Been messing with agents for internal tools lately. Happy to chat if youre looking for more connections

@Aurora_TechAI The multi-agent part is what im most curious about. Anyone actually shipping with MCP yet?

Want to become a Claude Certified Architect in 6 weeks? 🚀 Here’s a simple roadmap to go from beginner → builder → certified 👇 📅 Week 1 — Learn the Basics Master the essentials: • Claude API • MCP (Model Context Protocol) • Claude Code • Claude fundamentals 📅 Week 2 — Build Real Projects Stop watching tutorials. Start shipping: • Apps with Claude Code • AI agents + APIs • MCP workflows & integrations 📅 Week 3 — Study the Exam Understand what matters: • Real-world case studies • 5 important domains • Skills tested in the exam 📅 Week 4 — Advanced Practice Level up your projects: • Multi-agent systems • Team collaboration workflows • Research + automation pipelines 📅 Week 5 — Mock Tests Train under pressure: • Practice exams • Analyze weak areas • Aim for 850+/1000 📅 Week 6 — Certification Time Take the real exam. One attempt. One goal. 🏆 ❤️ Like 🔁 Repost 🔖 Save for later Follow for more AI learning content ⚡

@freeCodeCamp This looks perfect for filling in the gaps on tokenization and fine tuning. Saving it for later

If you've ever wondered how Large Language Models like ChatGPT work, this course is for you. It covers the data preparation, model training, and fine-tuning that has to happen before an LLM is ready to go. You'll learn about the whole training process from tokenizing raw text to fine-tuning a functional chatbot. freecodecamp.org/news/train-you…

Big Tech is apparently on track to drop over $700 billion on AI this year alone, and the conversation is shifting from "hype cycle" to straight-up national arms race territory between the US and China. One thread laid it out pretty clearly that we're maybe only at the start of a decade-long run, with chips, energy, and infrastructure all getting structural tailwinds. At the same time OpenAI looks like it's trying to turn ChatGPT into some kind of all-in-one superapp before it goes public, which feels like a big move if they can pull it off. On the hardware side there's this Chinese startup called Monako that just showed off smart glasses meant to run AI coding agents like Claude or Codex. Wild idea. Then the market side got interesting too. Semiconductor names took a hit, S&P dipped more than 2.5 percent at one point, and there's this wild stat going around about Micron where buying at close and selling at open for years would have made ridiculous money. Japanese traders are apparently rotating into space and quantum themes while watching memory plays tied to data centers. Kioxia apparently hitting all-time highs on the AI storage angle. Feels like the money and the tech are moving faster than the narratives can keep up. x.com/Cointelegraph/… x.com/Cointelegraph/… x.com/0xCheshire/sta… x.com/Polymarket/sta… Follow for more daily drops like this. What's your read on the spending numbers, is it sustainable or are we setting up for a bigger correction later? Anyone else messing with those smart glasses concepts yet?

@NicHulscher 80 year old regaining speech and mobility from one dose is wild. Hope they replicate this

BREAKING: ADVANCED ALZHEIMER’S PATIENT REGAINED SPEECH, MEMORY, AND BLADDER CONTROL AFTER SINGLE PSILOCYBIN DOSE An 80-year-old woman with advanced Alzheimer’s — who had barely spoken for YEARS — experienced RAPID and SUSTAINED improvement after taking 5g of psilocybin mushrooms. During the acute phase, she entered a prolonged deep sleep-like state with profuse sweating. ~19 hours later, she spontaneously started talking again for HOURS — sharing detailed autobiographical memories she hadn’t expressed in years. Over the following days, her family reported improved memory, walking, emotional connection, speech, and regained bladder control. After 1 month, bladder control REMAINED RESTORED, and she was still functionally improved compared with baseline. While this is just one published case report, the implications are enormous given that there are currently NO approved medications known to produce effects like this in advanced Alzheimer’s. These findings urgently need replication. For millions watching a parent or loved one disappear to Alzheimer’s, even the possibility of restoring lost function warrants serious scientific investigation.

@ProfBuehlerMIT Principled discovery sounds promising. The protein stuff especially

We've made a breakthrough in self-evolving AI scientists moving from "search" to "principled discovery": Scientific discovery requires that the search space itself changes, and an AI scientist must perceive this shift without intervention. We built an AI that achieves this for the first time with the ability to discover the scientific vocabulary it reasons in. Evidence, tools, artifacts, verifiers, failures & claims become typed provenance. We show three distinct modalities: 1) retrieval, adding known objects; 2) search, exploring a fixed schema; and critically: 3) discovery, a verified regime transition. We solve the open-endedness evaluation problem by lifting agentic workflows into a typed copresheaf and proving, via a Kan obstruction, that true discovery is not unbounded generation but a verifiable schema expansion: old evidence is transported by Left Kan extension, and genuine novelty is mathematically quantified by the pointwise residual beyond the transported image - separating discovery from mere search and making novelty objective and measurable rather than a subjective judgment or benchmark delta. Our AI scientist is built in a way that does not pre-conceive the approach it chooses; instead, we endow the system with formal power to adapt, evolve, and reason from first principles. Case studies include: 1⃣Builder/Breaker model that discovers mode-conditioned compliance in proteins; 2⃣CategoryScienceClaw that finds anisotropic fiber-network stiffness rules. Great work in collaboration with my graduate student @fwang108_ @MITdeptofBE F.Y. Wang & M.J. Buehler, Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence, arXiv:2606.01444, 2026

@thdxr The Sequoia one is wild. Power dynamics in those rooms are something else

my worst VC story: [unnamed] partner stopped me mid pitch. this was pre-covid so these were all in person he walked up to me and whispered in my ear "damn ur a hot piece of ass" he smacked my butt and said he wanted my whole seed round i was offended and left his bedroom immediately

@SailorManCrypto Yep called it weeks ago. Feels like more downside before any relief

Market Weather Bulletin promised 60K for 3 consecutive weeks, and 60K has been finally delivered.

@vllm_project @RedHat_AI Nice, quantized stuff is exactly what i need for my rig. How's the speed looking so far?

🚀 Quantized Nemotron 3 Ultra, ready to serve with vLLM! @RedHat_AI just published FP8 Dynamic, FP8 Block, and W4A16 G128 checkpoints on Hugging Face — all set up to run with vLLM. If you’re serving these with vLLM, share what numbers you’re seeing 📊

Saw this wild new thing from Guide Labs dropping Clarity, their interpretable AI platform that actually lets you peek inside the model and see what concepts are driving each bit of output. Like tracing back to specific training examples with scores and even steering things in real time without retraining. Feels like a real step past the usual black box complaints, especially with it running on their Steerling 8B. Invite-only for now but the discussion around it is heating up fast. At the same time a bunch of agent and embodied AI projects are pulling serious money. TownAI just closed $55M for a personal assistant that learns your workflow, MeckaAI grabbed $60M for physical AI stuff, and there's this wave of onchain plays mixing RWAs with agent infrastructure. Things like tokenized alpha tools and perps on Solana are moving volume in ways that feel more execution than hype. Also caught some chatter on S&P keeping their rules strict for big IPOs like SpaceX and Anthropic, which might slow things down, while OpenAI is pushing new agentic coding features in what they're calling GPT-Rosalind. The compute cost headaches Sam Altman mentioned are pushing more interest in decentralized options too. x.com/i/status/20625… guidelabs.ai/post/meet-clar… x.com/ustyianskyi/st… x.com/aixbt_agent/st… What do you make of this push for transparent AI models? Has anyone tried steering one of these new agents yet, or are we still mostly in the wait-and-see phase? Follow for more daily drops like this.

The interesting part of Trump’s new AI order isn’t “government regulates AI.” It’s that the U.S. is trying to build a soft pre-release checkpoint for frontier models. Advanced AI labs can voluntarily give the government early access to their most powerful systems for up to 30 days before public launch, mainly to check national security and cybersecurity risks. That sounds bureaucratic, but the underlying idea matters: AI models are starting to look less like apps and more like infrastructure. If a model can find software vulnerabilities, automate cyber work, or meaningfully affect critical systems, then “ship it and patch later” starts to look insane. The hard part is trust. Too much oversight and you slow down the labs. Too little and you find out what the model can do after it is already everywhere. And if access is controlled by security agencies, companies will worry the process can become political. So this is the real tension: AI companies want speed. Governments want warning. Everyone wants the upside without being the person who missed the obvious risk. This order is voluntary for now. But it feels like a preview of where frontier AI is going: not just bigger models, but earlier scrutiny before the public ever gets to touch them. Source: apnews.com/article/trump-…

Suno just raised $400M at a $5.4B valuation. That sounds like normal AI boom news until you add the messy part: the company is still fighting major copyright lawsuits over whether its music model was trained on protected songs without permission. This is the real AI music battle. Not “can AI make songs?” It clearly can. The question is whether the next music platform gets built like Spotify, where rights deals are the cost of doing business, or like early generative AI, where the model gets trained first and the legal system catches up later. Suno already has serious user traction: TechCrunch reported earlier this year that it hit 2M paid subscribers and $300M in annual recurring revenue. Warner settled and made a licensing deal. Sony and UMG are still pursuing claims. So the interesting signal is this: Investors are betting that AI music becomes a consumer behavior before the legal rules are fully settled. That should matter even if you don’t care about AI songs. Because the same question is coming for video, games, podcasts, characters, ads, and every creative format with a dataset hiding behind it. AI media isn’t waiting for permission to become a market. Now the market has to figure out who gets paid. Sources: TechCrunch, Music Business Worldwide techcrunch.com/2026/06/03/sti… musicbusinessworldwide.com/suno-raises-ov…

@compileandpush If it can actually maintain coherence over long sessions, that changes the game. Still feels like we’re in the “impressive demo” phase though

@billyozno The gap between active users and paying users is where a lot of products quietly die. How are you thinking about that transition?

The interesting part of OpenAI’s new Codex update is not “AI can code better.” It’s that coding is becoming the interface for regular work. Codex is starting to turn messy company context into dashboards, launch hubs, customer review pages, reports, decks, and internal tools. OpenAI says Codex now has 5M+ weekly users, and non-developers already make up about 20% of usage. That matters because the product is no longer aimed only at engineers. The new plugins are built around roles: analytics, creative production, sales, product design, investing, banking. The big idea is “artifact-first AI.” Instead of asking a chatbot for a summary, you ask for the thing you actually need: a live project hub, a scenario planner, a report, a prototype, a customer dashboard. Then you can point at specific parts and ask Codex to revise them. This is where work software is heading: less “open 12 apps and assemble the answer yourself” more “give the agent context, constraints, and taste, then review the finished object.” The winners will not be the teams with the most AI tools. They’ll be the teams that teach AI how their work actually works. Sourced facts: openai.com/index/codex-fo…