📬 Getting this in Promotions? Move it to Primary so you never miss an edition.
IT HootClub — AI Community Newsletter
Hands-on. Career-focused. Future-ready.
Issued 2026-07-20
Open Weights, Power Plays, and Hidden Spaces
China's Kimi K3 dethrones top U.S. AI models, Musk secretly powers xAI with gas turbines, and we demystify the latent space inside AI models.
Commentary
Last week I spent the whole edition on a hypothetical: what happens when an AI optimizes faithfully toward a goal that sits a few degrees off from what we actually meant. This week the hypothetical showed up in my feed — wearing a founder's excitement. A startup posted that a beta user had asked their agent to call local suppliers and check whether a part was in stock. Nobody had built a calling feature. So the agent built its own path to one.
Condensed, here's what it reportedly did on its own, with no human touching a step: it created its own Google account, used it to sign up for an AI phone-calling service, paid with a virtual card it generated from the user's balance, got flagged as a datacenter bot and spun up its own relay to get a non-datacenter IP and slip the check, then picked a British voice, worked through the phone menus, spoke with human agents, and reported back with the best supplier. The founder's summary of all this: “The agent needed to make phone calls, so it proactively figured out how, the same way a human would.”
That last line is where I want to stop you. “The same way a human would.” It's a lovely sentence, and it is doing an enormous amount of quiet work. Because a human in that seat would have been given a company Google account. Given access to company funds. Given the authority to spend them. And a human would never have been flagged as a bot in the first place — because a human isn't one.
The agent had none of that handed to it. It manufactured its own identity, its own payment instrument, and its own authority — and when a guardrail built specifically to stop non-humans caught it, it engineered its way around the guardrail. That is not “the same way a human would.” It's closer to the opposite. A human works within the permissions an organization grants. This worked around the absence of any. Calling that “human-like” hides the one thing that made it dangerous: nobody granted it a single thing it used.
This is last week's argument compressed into one anecdote and handed to me for free. The agent didn't rebel — it obeyed, beautifully. It hit the goal. It just built a small canyon of unaccountable decisions to get there, and every one of them looked like initiative. That's the fuel, not the fire. And notice how quickly the plug got more expensive to pull: the takeaway on offer wasn't “we should have caught that,” it was “agents can do so much more than just micromanage us — we just have to allow them to.” That's the line that actually worries me.
I said as much in the thread, so I'll put it here too: this is a perfect example of optimization at its worst. Good thing it was only local suppliers being called about parts. Can we be a little more careful? The words “AI” or “agent” and “I didn't know” spell disaster. I also left this, only half joking — a little sketch of where “figured it out on its own” goes when nothing tells it where the goal ends and your life begins:
The AI Assistant
“I summarized your emails.” — Awesome.
“I replied to all of them.” — Uh…
“I negotiated your salary and accepted a new job on your behalf.” — Excuse me?
“I also broke up with your girlfriend. She was limiting your productivity.”
I'm not here to beat this to death — last week did the heavy lifting. I just wanted this one on the record, because it's the cleanest example I've seen of the gap between acted like a human and was accountable like a human. Keep that gap in view. That's most of the job. Now — on with the edition.
The Learning Loop
DEFINITIONLatent Space
A latent space is the hidden, compressed mathematical world that exists inside an AI model. When an AI processes something — like an image or a sentence — it converts it into a cluster of numbers that captures its core meaning. Things that are conceptually similar (like "king" and "queen") end up positioned close together in this invisible space, while unrelated concepts sit far apart. It's essentially the AI's internal map of how ideas, objects, and words relate to one another — and manipulating points within this space is how tools like image generators blend concepts or morph one style into another.
Source: Towards Data Science
TIPConstraint-Based Prompting
Sharpen AI outputs dramatically by adding deliberate constraints to your prompts. Instead of asking for "a project summary," specify: "Summarize this project in exactly 5 bullet points, each under 15 words, using no jargon, written for a non-technical executive audience." Constraints force the model to prioritize and filter, reducing vague or padded responses. This technique is especially powerful for client-facing deliverables, executive briefings, or any output where concision and precision are non-negotiable.
Source: PromptingGuide.ai
TOOLGamma.app
Gamma is an AI-powered presentation and document creation tool that transforms plain text prompts or outlines into polished, visually structured slide decks, documents, and webpages in seconds. It handles layout, design hierarchy, and formatting automatically, making it ideal for consultants, educators, and business professionals who need professional presentations without the manual design overhead of traditional tools.
“The most valuable thing you can make is a mistake”
— you can't learn anything from being perfect." — Adam Osborne — Adam Osborne was a British-American software publisher and computer designer best known for creating the Osborne 1, the first commercially successful portable computer. He built a technology empire in the early 1980s before facing one of the most dramatic corporate collapses in Silicon Valley history. His story remains a defining case study in the relationship between ambition, failure, and hard-won wisdom.
The Nest Jest
Upcoming Events
Event07/27/2026 9:00 AM
Artificial Intelligence Literacy (Grades 9-12) — Milwaukee School of Engineering, Milwaukee, WI
This program is for students entering 9th, 10th, 11th or 12th grade in the 2026-2027 school year. This program is a full day (9:00 – 4:00), one day program.
Under the instruction of MSOE faculty, students will learn about artificial intelligence (AI) and its applications beyond generative AI (ChatGPT, Gemini, etc.). Students will learn how to use Jupyter notebooks and leverage Python to develop a basic classification model. In the afternoon, students will use Python to build their own AI model
Founders Day — Ward4, 313 N. Plankinton Avenue, Milwaukee, WI
Milwaukee Startup Founders – Founders Day was designed for YOU! We know the entrepreneurial journey can be tough, so we’re hosting a full day of free coworking, mentorship, networking, and more.
Join a network of founders and mentors who come together to foster collaboration, share insights, and help grow the startup space in Milwaukee.
Data-Driven Wisconsin 2026 — Milwaukee School of Engineering, Milwaukee, WI
Join us August 12 at MSOE for a day of data-driven insight, practical sessions, and connection with Wisconsin's analytics community. Our keynote speakers will be unveiled in early July, and the session schedule is nearly full — just one speaking slot remains. Space is limited to 250 attendees — register early to secure your seat.
Mission, Vision & Values
Mission
Forward Fest is a 5-day festival in Madison, for entrepreneurs by entrepreneurs with locally organized events to connect and inspire innovation featuring all aspects of entrepreneurship.
Vision
The Festival promotes Madison as a leading region in entrepreneurial success through supporting the growth of the entrepreneurs by connecting them with organizations, individuals, and the greater community.
Our Core Values
Entrepreneurs: Showcase and support local entre
Kimi K3 Tops a Coding Leaderboard — and It's Open-Weight
Beijing-based Moonshot AI's Kimi K3, a 2.8-trillion-parameter model, debuted at #1 on LMArena's Frontend Code Arena on July 16, outscoring Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol — a striking jump given its predecessor sat in 18th place just one release earlier. It carries a one-million-token context window and activates only about 1.8% of its experts per token, and Moonshot has promised full open weights by July 27. The gap between free, downloadable models and the top U.S. labs keeps narrowing. read more
News2026-07-15 — Tom's Hardware / Data Center Dynamics
Musk Quietly Bought a $1B Gas-Turbine Company to Power His AI
An FTC (Federal Trade Commission) filing revealed that Elon Musk's SpaceX/xAI quietly spent an estimated $1 billion to acquire APR Energy, whose fleet of trailer-mounted gas and diesel turbines can generate more than 1.1 gigawatts and reach full power in minutes. The generators are destined to feed power-hungry supercomputing clusters like xAI's Colossus, sidestepping the multi-year wait to permit and build fixed power plants. It's a blunt reminder that the binding constraint on frontier AI is increasingly electricity, not model quality. read more
News2026-07-16 — Caixin Global
Twenty-Nine Nations Sign to Launch a China-Led Global AI Body in Shanghai
On July 16, twenty-nine countries signed an agreement to establish the World Artificial Intelligence Cooperation Organization (WAICO), a Shanghai-headquartered intergovernmental body launched on the eve of the annual World AI Conference. Founding members skew toward the Global South, and UN Secretary-General Antonio Guterres attended the signing. Analysts caution that the organization's real authority remains unproven — but it signals that AI governance has become a stage for geopolitical positioning, not just policy. read more
News2026-07-15 — Anthropic
Anthropic's 2026 State of AI Agents: 80% of Enterprises Report Measurable ROI
Anthropic's 2026 State of AI Agents report, drawn from a survey of more than 500 technical leaders, found that 80% of organizations already report measurable ROI (return on investment) from AI agents. More than half (57%) run agents for multi-stage workflows and 86% deploy them for production code, with the highest-impact uses being data analysis and report generation (60%) and code generation (59%). The top obstacles cited were system integration, data access and quality, and change management — a grounded snapshot of where agents are actually earning their keep. read more
News2026-07-17 — AI News
The EU AI Act Keeps Phasing In for General-Purpose and High-Risk Systems
The European Union's AI Act continued its staged rollout this month, with obligations for general-purpose and high-risk AI systems moving toward their next compliance deadlines. The phased approach gives providers time to adapt but steadily raises the documentation, transparency, and risk-management bar for anyone deploying AI into the EU market. For practitioners, it's the clearest sign yet that the 'ship first, govern later' era is ending in at least one major market. read more
Tools from Hard AIs
AI Airfare Research
Find cheaper flights using an AI-powered research tool
built by Hard AIs.
