Quick recap: my name is Liberato Aguilar, founder of Aguilabs. A lot has happened since the last issue, in both local and national AI news.
If you want the full context (3 min read), check out our last issue.
Glad you're still here. Let's get started.
In the last issue I mentioned Anthropic had released their most powerful model yet. The model is good. REALLY good. Honestly the most capable coding model I have tried, and a clear step up from Opus 4.8 and GPT 5.5. I got a full three days with it before the US government stepped in and forced Anthropic to shut it down.
So what happened?
Fable 5 is the locked-down sibling of Mythos, a private model only available to select Anthropic enterprise partners. Fable ships with heavy guardrails that refuse any cybersecurity and biology prompts. The government says it became aware of a "jailbreak" that slipped past those guards.
Per Anthropic's own statement, the jailbreak basically came down to asking the model to read a codebase and fix software flaws, and it only turned up a few minor bugs that were already known. Anthropic also says other public models like OpenAI's GPT 5.5 can do the same thing.
The government is erring on the side of caution though, so they issued an export control directive. The actual order is: no foreign national can use Fable or Mythos, inside or outside the US. Anthropic cannot verify the nationality of every single user, so the only way to comply was to pull the model for everyone. Even some of their own employees and researchers.
The timing is rough. Two days before the ban, CEO Dario Amodei published an essay arguing the government should be able to block unsafe AI models, modeled on how the FAA handles aviation safety. Then the government did exactly that. Anthropic's response is that this particular action does not follow the fair, transparent, fact-based process Dario asked for. He basically handed them the playbook and did not love how they ran it.
This is not the first time Anthropic has clashed with the US government. A few months ago the DoW labeled them a national supply chain risk. I get the worry. Mythos-class models are good enough at finding software vulnerabilities that Anthropic itself flagged them as a national security risk, and any US adversary would jump at the chance to use them.
This is the same problem I dug into in my StuxSwarm paper, where AI agents hunt for exploits faster than people can patch them.
The bigger question the ban opens up is: should the government get to decide who can and cannot use a given AI model? Has the age of AI passports begun?
While the US took its first real swing at AI regulation, China went the other direction.
Chinese labs have become the front runners in open source AI, models released publicly so anyone can see how they were built, or build on top of them. They usually land just below Anthropic or OpenAI in performance, but they cost a fraction of the price, and for plenty of use cases that trade-off is worth it.
This week the lab Z.ai (formerly Zhipu) released GLM 5.2, an open weight model that lands within a few points of Opus 4.8 on coding and beats GPT 5.5 on others, at roughly a sixth of the cost.
From my personal initial testing, GLM 5.2 is a good model and I have my main assistant AI agent running on it now. It's great at handling quick tasks like logging expenses, managing invoices, and sending project updates to clients.
As models keep getting smarter, we will find out whether tighter US rules keep America ahead or hand the edge to China's open approach.
Elon Musk is officially the world's first trillionaire.
He crossed the line on June 12, the day SpaceX went public, and his net worth now sits north of a trillion dollars, more than the next several richest people on earth combined (CNBC).
The stock priced at $135, opened at $150, and ran as high as $225 within a few days. It has been a roller coaster since. As I write this it is trading around $156, still up about 15% from the IPO price, with a market cap near $2.06 trillion.
A few days after the IPO, SpaceXAI, which folded in Musk's AI company xAI earlier this year, agreed to buy Cursor, the AI coding platform, for $60 billion in stock. The deal still needs regulatory approval and is expected to close next quarter.
Cursor is a four year old company started by four MIT students. The buyout now makes each ~25-year-old founder worth $2.7 billion! Congrats to the Cursor team. I have used their platform for Aguilabs dev work myself.
In local news, El Paso City Council voted 5 to 3 against starting the process to cancel the city's incentive deal with Meta, so the data center going up in far Northeast El Paso keeps moving forward (KTSM).
In the last issue I went deep on this project and what it could mean for our community. If you are new here and want to understand how these data centers actually use water, here is my writeup.
So what does the vote mean for us?
The project continues. Meta now calls it a $10 billion investment and says it will bring about 300 permanent jobs plus up to 4,000 construction jobs at peak buildout.
One thing we should keep in mind is that the incentive agreement itself only requires Meta to keep 50 full time jobs. The city also recently voted to stop handing out financial incentives for any new data centers, so this project may be the last one of its kind here for a while.
This is where I shine a light on local partners around the border and support small businesses. We recently teamed up with 350° Delights to build an ecommerce shop for their custom baked-goods gifts.
Shopify or Square templated websites were not sufficient for our client's needs so we created a tailored-experience to best showcase what 350° Delights has to offer.
350° Delights makes delicious pastries for any occasion, from birthdays to baby showers to whatever you are celebrating. You can place an order for pickup or delivery at 350delights.com.
The World Cup is here, and with the US, Mexico, and Canada all hosting, you can feel the excitement everywhere!
If you have caught any of the group stage, you have probably noticed the new Referee Cam, a small headset camera that gives you the referee's point of view.
Lenovo developed it and runs a real time AI model that stabilizes the feed, cutting motion blur by up to 50% so the footage is actually watchable. It's live in all 104 matches, a World Cup first (Lenovo).
For the devs
Most matches have landed during work hours, which makes it hard to follow live scores.
I spend most of my day coding and running AI agents in the terminal, so I built a World Cup marquee for tmux that scrolls through group standings, upcoming matches, and live scores right where I work.
There is also a purchasable slot, so you can broadcast a message to everyone who has it installed.
The whole thing is open source! Check it out at superchat.aguilabs.com.
That wraps the second issue of Aguilabs Signal.
The AI space moves fast, and my job is to keep the local community in the loop. If you got something out of this one, please share it. And if you have questions about how Aguilabs can help your business adopt AI or build custom software, reach out anytime.
Coming up next issue:
• We are exploring the new agentic era of finance, looking to use AI agents to manage investments and outperform the S&P 500. More on that soon!
• Can LLMs create a better World Cup bracket than you? See how Claude's bracket has performed and whether we should trust AI with sports predictions.
As always, thank you for the support,
Liberato and the Aguilabs team 🦅
