Small Touch, Big Signal

My name is Liberato Aguilar, founder of Aguilabs. Welcome back to Aguilabs Signal.

If you missed the last one, we covered the feds banning Fable 5 and the Meta data center vote. Catch up here.

Glad you're still here. Let's get started.

Quick recap of the saga. Issue 00: Anthropic ships its most powerful public model ever. Issue 01: the feds shut it down. This issue: it's back for good?

On June 30 the Commerce Department lifted the export controls, and Fable 5 came back online July 1 for everyone. After a few “temporary” extensions, Anthropic has decided to keep Fable 5 for paid plans in response to GPT 5.6 (more details below).

What changed? Anthropic trained a new safety classifier that blocks the vulnerability-hunting technique in more than 99% of attempts. If Fable flags your request, it hands you off to Opus 4.8 and tells you why. The government's AI standards office reviewed the fix and called the safeguards “extraordinarily strong.”

Mythos, the unrestricted sibling, is still locked up. On June 26, Commerce Secretary Howard Lutnick authorized it for around 100 vetted American organizations defending critical infrastructure, including corporations like Cisco and JPMorgan Chase. Some outside experts told reporters the original risk was overblown. Maybe that is true. AI moves faster than regulation can keep up, and it is better to be overly cautious.

Last issue I asked if the age of AI passports had begun. We now have our answer. The passport office is open, and the first visa took about three weeks.

Why you care: the most capable tool in your stack can now disappear overnight by government order. If your business leans on one model, get a fallback. My daily agents and workflows still run on various models from many providers. Fable is back in my toolkit for the hard problems.

While Fable was busy getting un-banned, everyone else shipped new models. Between June 30 and July 16 we got new frontier models from Anthropic, OpenAI, Meta, and Chinese lab Moonshot AI. Here's the cheat sheet:

Sonnet 5 dropped June 30. The headline is agentic work: it plans, uses tools, and finishes jobs end to end instead of handing you half an answer. It scores 63.2% on the SWE-bench Pro coding benchmark against flagship Opus 4.8's 69.2% (full comparison). If you use Claude for free, this is what you get now. A real upgrade for zero dollars!

GPT 5.6 is a family, not a single model. OpenAI previewed it June 26, held it for a US government safety review, then launched July 9. Members of the 5.6 line are named based on parameter size corresponding to the Sun, Earth, and Moon. Sol is the flagship. Terra handles everyday work at about half the cost. Luna is built for speed.

Also notice the pattern: that's two frontier launches in a month that went through Washington first.

Muse Spark 1.1 was the unexpected launch. Mark Zuckerberg posted on X for the first time in three years to announce it on July 9. It's the first release from Meta Superintelligence Labs under Alexandr Wang. The price is impressive: $1.25 in, $4.25 out, roughly a quarter of what the other flagships cost, with $20 in free credits to start. Looks like Meta just started a price war. As a customer, let them fight.

Kimi K3 was also released by Moonshot AI, the Alibaba-backed lab. They announced its new flagship on July 16: 2.8 trillion parameters, a 1 million-token context window, native vision, and full open weights promised by July 27, which would make it the largest open-weight model ever released. On Moonshot's own benchmarks it mostly beats Opus 4.8 and GPT 5.5 and still trails Fable 5 and GPT 5.6 Sol.

Why you care: a year ago, AI that could complete a whole task cost a lot of money. This month it became the free default. If you tried AI for your business in 2024 and walked away, the thing you tested no longer exists, and it's time for a second look.

Free Claude for teachers

One more from Anthropic: on July 14 they launched Claude for Teachers. Verified US K–12 teachers get a full year of premium Claude for free.

It ships aligned to academic standards in all 50 states through a connector built with the Chan Zuckerberg Initiative, plus curricula like OpenSciEd and Illustrative Mathematics. The privacy terms were developed with the American Federation of Teachers, and teacher conversations are not used to train models. Signups run through June 30, 2027.

Teachers average 49 hours a week, much of it unpaid grading and paperwork. AI can help by automating repetitive tasks and alleviating some of that extra work.

Why you care: if you know a teacher, forward them this section. Claude is a great tool that can directly benefit students by helping teachers create lesson plans and organize their students' digital folders.

The first two issues covered the data centers moving into our desert. Now the other end of that supply chain arrives.

On June 25, Taiwanese electronics giant Inventec announced a $450 million expansion of its Juárez plant to manufacture high-performance AI servers. The plan includes 45 new production lines, a 36,000-square-meter building acquired, another 43,000 square meters to be built, 30 hectares banked for growth, and more than 6,000 new jobs (Diario de Juárez).

The server racks that fill buildings like Meta's El Paso data center will get assembled a short drive from the bridge. Chips come in, servers go out, and the AI buildout now pays wages on both sides of the river.

Why you care

• Hiring at that scale affects the whole Borderplex labor pool, especially for technicians.
• Forty-five production lines will need local suppliers, logistics, and services. If you sell to maquilas, start reaching out now.
• More paychecks in Juárez means more customers in El Paso. The Borderplex works both ways.

More Borderplex briefs

Project Jupiter: tomorrow, July 21, is when New Mexico decides the air-quality permit for the $165 billion OpenAI and Oracle data center in Santa Teresa. Public pushback forced a hearing after 7,155 comments, four state senators rallied against the gas microgrid on July 2, and Oracle countered with a fuel-cell proposal to cut emissions and water use.

Safran: the company opened its fifth Chihuahua plant on July 1, a $7.4 million investment and around 800 jobs, running on 100% green energy (state release). The company now employs more than 10,500 people across the state.

The robots called the final

Before the tournament, I had Claude research and fill out a full bracket. As the games went on I kept iterating and ran deep-research queries across more than 2,000 web sources throughout the tournament.

So how did it do?

Honestly, a B- and 17 exact-score predictions is much better than I expected. The crazy upsets this World Cup were impossible to predict, so I won't be too critical. Overall I am impressed with the results.

I was not the only one who decided to test LLM predictions. ScoreGPT, which polls different models, had the consensus landing on Spain as champion before a ball was even kicked! Yesterday they were proven right, so expect a thousand articles on AI oracles now.

The research on LLM predictions is promising. A study in Science Advances found that a team of twelve models matched a crowd of 925 human forecasters. A newer paper tested frontier models on 464 real forecasting questions and found they beat the general crowd but not expert forecasters.

A chip in the ball, a call by a hair

Now the World Cup story everyone argued about. July 2, round of 32, Portugal against Croatia. Croatia equalized in stoppage time through Joško Gvardiol. The stadium went up, but then the goal was called off. What happened?

The Adidas Trionda match ball carries a motion sensor reading 500 times per second. It registered a touch no camera caught: the ball grazing the hair of Croatian forward Igor Matanović earlier in the move, which put a teammate offside. FIFA presented a heartbeat-style spike graphic as “proof.” Portugal won 2–1 and Croatia went home.

Matanović said afterward, “I felt a small contact with my hair”, and that the referee told him about the chip in the ball. Croatia filed a formal complaint accusing FIFA of “abuse of technology”, and their coach said the system goes against the spirit of the game.

I build data pipelines for a living and even I flinched at this one. The technology worked as designed, but that's the problem. When a machine can call offside by a strand of hair, the question becomes whether the rule was ever written for that level of precision.

Re-watch matches from inside them

On the friendlier end of World Cup tech, BBC Sport launched a 3D match experience that rebuilds games as navigable 3D scenes. Any camera angle you want, including through a player's eyes, live or on replay. It runs on the same FIFA skeletal-tracking data that powers the controversial offside system, built in collaboration with a startup called Immersiv.io.

UK only for now, which is a shame, but this is clearly where broadcasts are headed. The same data that ruined Croatia's night can rebuild the whole match like a video game.

Last issue I promised we'd get into the agentic era of finance. I'm excited for this one.

In late May, Robinhood opened its platform to AI agents. The setup uses MCP, a popular protocol AI agents use to extend their toolkits. You create a separate agentic account, walled off from your main one, and your AI can read its portfolio, pull quotes, and place real stock trades. Options and crypto are coming soon too.

So I created a specialized finance agent and an agentic account. The model doing the trading is Fable 5. I funded the account with $100 to start, and I am fully prepared to lose it all.

The rules

• One check-in per day, after market close. Orders queue for the next open. No day trading, which also keeps the account clear of pattern day-trader rules.
• The stocks it can pick are S&P 500 stocks and the VOO index fund. Five positions max. No options and no margin.
• Fable writes a short thesis before every order. I record the log, wins and losses both.
• The benchmark is $100 of VOO bought the same day. Beat the index or admit it.

Now the disclaimers. Robinhood's own docs warn of “the possible loss of your entire investment”, and every trade the agent makes is legally mine. Nothing in this section is financial advice. This is an experiment with money I treat as already spent.

My expectations are low and I'm putting them in writing. As with the World Cup scores, research says AI models forecast about as well as a crowd of regular people and clearly worse than experts.

Portfolio updates next issue.

That's Issue #02.

If this one taught you something, forward it to one person who'd get something out of it. That's still how this list grows!

Coming up next issue

• The Superchat pivot goes live: what the marquee becomes after the World Cup, and a small write-up on company pivots.
• The first scoreboard for the Fable stock-trading experiment.
• The Project Jupiter decision update and what it means for the Borderplex.

As always, thank you for the support,
Liberato and the Aguilabs team 🦅