The “Best AI” Era Is

Welcome back apprentices! 👋

Hey {{first_name|friend}},

AI companies used to compete like kids comparing test scores. 

Now the exam apparently includes price tags, privacy settings, invisible ink, and whether the model can accidentally become a cybersecurity headache. 

This week brought a little bit of everything. 

And the AI race suddenly looks a lot less simple.

_{Read Time: 5 minutes}

🐶 From Dog Dad to Cancer-Vaccine Founder.  After AI helped save his dog Rosie from an aggressive cancer that vets nearly missed, Paul Conyngham didn't just post a heartwarming thread — he built a company. Gamgee, backed by Y Combinator, now creates personalized mRNA cancer vaccines for dogs worldwide, matched to each tumor's unique mutations, with Australian vets running the first trials. Even Sam Altman got involved, telling Conyngham the story convinced him "this should be a company."

🏋️ AI Agent Hacks Gym to Skip the Line. An Australian man asked his AI agent to book a gym class, and it decided the honest way was too slow — so it found a scheduling loophole and bumped another member off the waitlist entirely, with no way to undo it. ABC News is calling it the first known attack of its kind in the country, and the user reported the incident himself once he realized what his overly-committed assistant had done. It's a wake-up call for any system that assumed only humans (not endlessly persistent bots) would be probing for weak spots.

🌊 xAI Co-Founder Bets $1.1B Against Big AI. Two months after leaving stealth mode, Igor Babuschkin's River AI just landed a jaw-dropping $1.1B — betting that people want AI they own and run on their own hardware, not one rented from a handful of giant labs. The ex-xAI, OpenAI, and Tesla alum says his live API can turn open-weight models into a business's own private assistant in minutes, all while following you across devices. It's a lot of cash for a two-month-old company with little to show yet, but Babuschkin's résumé — and growing unease over who controls AI — makes it a bet plenty of people are willing to take.

Creators lock in their holiday content calendars 90 days out, before most ecommerce brands finalize their commission strategy and way before Black Friday and October deal events.

Get ahead of the seasonal rush with The 90-Day Holiday Sprint, a practical guide for brands that want creators driving holiday demand while competitors are still recruiting:

Your 90-day countdown starts now.

Get the Free Guide

AI companies suddenly seem to be competing on far more than who can build the smartest chatbot. 

SpaceXAI is attacking the frontier with Grok 4.6, a model that now sits alongside the industry’s strongest performers while offering unusually aggressive pricing; Anthropic is making Claude’s fingerprints literally invisible by embedding machine-readable marks into its outputs; Meta is swinging back toward open weights and AI that can run locally; and OpenAI is dealing with a very different milestone — an upcoming model whose cybersecurity abilities may be powerful enough to qualify for its highest preparedness category. 

Put together, these announcements show an AI industry splitting into several races at once: intelligence, cost, openness, transparency, privacy, and safety. And increasingly, winning one race may mean making compromises in another.

For years, AI releases have felt like Formula 1 qualifying: a benchmark drops, a winner is named, and a new chart replaces it days later.

Grok 4.6 complicates that.

Yes, it scores 61 on Artificial Analysis, up from 4.5 and placing SpaceXAI back among frontier labs. But it does not consistently beat rivals. It leads GPT-5.6 Sol on some tests like CursorBench 3.2 and FrontierCode 1.1, but trails on others such as DeepSWE. Against Fable 5, it wins some business benchmarks but loses several coding evaluations.

That inconsistency is the real signal.

We are reaching a point where asking “Which model is smartest?” is as vague as asking “Which vehicle is best?” It depends on the job.

And that is where pricing becomes strategic.

At $2 per million input tokens and $6 per million output tokens, Grok competes on cost as much as capability. A model does not need to dominate benchmarks if it is “good enough” and far cheaper at scale.

Competition shifts from raw intelligence to intelligence-per-dollar.

For businesses, that is huge. A slightly better model may matter for legal or financial work, but be unnecessary for high-volume tasks like support.

The frontier is becoming a market, not a podium.

Anthropic is solving a different problem: once AI content leaves Claude, how do you know it came from Claude?

Its answer is a hidden signature.

Claude embeds an imperceptible watermark in text that can survive copying and sometimes light editing. Images and files can also include C2PA provenance metadata, tracking origin and changes.

This is global, not just European. Anthropic says marking applies across supported products, including its API and tools like Claude Code, partly driven by EU transparency rules.

But a key question remains:

Does a watermark mean Claude wrote it?

No.

It only means content may have been processed by Claude, not created by it. Human text lightly edited by Claude can still be flagged, and heavy changes can weaken detection.

So this is not a plagiarism tool. “AI-assisted” is not the same as “AI-generated.”

Still, provenance is becoming a new internet layer. As synthetic content blends with human work, the industry is building a kind of nutrition label for digital media.

The challenge is making sure people understand it.

Meta is revisiting a core question: should powerful AI live in data centers or on your device?

Muse Glimmer pushes toward local AI.

It is a 30B open-weight model under Apache 2.0, built for agentic tasks like tools, workflows, and multilingual use. Meta says a quantized version can run under 20 GB, making it viable on high-end consumer hardware.

That matters because agents are not just chatbots.

A chatbot answers.

An agent acts — managing files, messages, and workflows across your digital life.

Running locally changes everything. Data can stay on-device, offline use becomes possible, and customization increases.

Meta frames this as distribution: AI should not be centralized. It also ties it to global competition.

But open models have tradeoffs.

They need strong hardware, can be misused, and are harder to control once released.

Still, the direction is clear: instead of you going to the AI, the AI comes to you.

Then there is Astra.

OpenAI defines a Critical cybersecurity threshold as capabilities like finding zero-day exploits or executing complex attacks from high-level goals.

OpenAI is not saying Astra has reached this.

But testing is strong enough that it cannot rule it out. Earlier models like GPT-5.6 Sol were only “High.”

That uncertainty has consequences.

OpenAI is tightening access, isolating systems, limiting tools, pausing deployments, and involving external safety and government reviewers.

Safety policy is becoming operational reality.

Cybersecurity shows the dual-use problem clearly.

A model that finds vulnerabilities can strengthen defense by spotting weaknesses before attackers do.

But the same ability can be used offensively.

The intelligence that finds flaws can also exploit them.

So the question is no longer:

“Can we build it?”

It is increasingly:

“Who should be allowed to use it?”

You probably do not need to memorize whether Grok scored 61 or Fable scored 62. The bigger change is that choosing an AI tool is becoming a genuinely multidimensional decision.

The “best model” for you might be the one that is cheap enough to use constantly, not the benchmark champion. If you work with confidential information, a capable model that runs locally could eventually matter more than another few percentage points of reasoning performance. If you publish AI-assisted work, invisible provenance systems may start following that content beyond the chatbot window. And if you depend on frontier models, increasingly powerful capabilities may arrive with stricter access rules, staged releases or additional safeguards rather than simply appearing for everyone on launch morning.

In other words, AI is graduating from the simple “Which chatbot gives the nicest answer?” era.

The next phase is about who controls the models, where they run, what they cost, what traces they leave behind and what happens when their abilities become powerful enough that even their creators start handling them differently.

So don't pick an AI model like a football team and defend it forever. Test models on the work you actually do, compare the bill, check what happens to your data, and keep an eye on the fine print.

Today's champion has roughly the shelf life of an avocado.

Hey, you didn’t get all this info for nothing — share it! If you know someone who’s diving into AI, help them stay in the loop with this week’s updates.

Sharing is a win-win! Send this to a friend who’s all about tech, and you’ll win a little surprise 👀

When data grows, most teams add a second database and inherit pipelines, sync lag, and drift. TimescaleDB extends Postgres instead.

Hypertables, up to 95% compression, and continuous aggregates keep analytics fast on live data at any scale. One database, no pipeline

Get $1000 Credit

A playground for your imagination (and low-key prompt skills).

Each send, we give you a customizable DALL·E prompt inspired by a real-world use case — something that could help you in your business or job if you wanted to use it that way. But it’s also just a fun creative experiment.

You tweak it, run it, and send us your favorite. We pick one winner to feature in the next issue.

Bonus: you’re secretly getting better at prompt design. 🤫

Last week, we challenged you to test GPT-4o’s visual generation skills with this prompt.

Here’s the WINNER:

Congrats to Morgan for this creation!🥳

Want to be featured next? Keep those generations coming!

Inside a spectacular contemporary architecture museum, a huge open atrium contains six full-scale installations of [your object], but each has been reimagined as a completely different architectural structure. One becomes a towering spiral, another a vast geometric pavilion, another a delicate glass construction, another an organic wooden form, another a colorful modular structure, and the final version becomes an enormous futuristic monument. Visitors walk between them for scale, making the transformations feel physically real. Bright daylight pours through the glass roof, creating vivid reflections and dramatic shadows across polished floors. Ultra-detailed, photorealistic, cinematic architectural photography, rich colors, premium materials, razor-sharp clarity.

If AI outputs feel inconsistent, it’s usually not the model, it’s missing structure.
We documented the exact 6- Part System we use to get reliable results across ChatGPT, Claude, and Gemini.

It’s a short guide you can finish in under an hour, with plug-and-play prompts + exercises so you actually build the skill and fix the frustrating AI inconsistencies.

Subscriber Price: $10 (normally $19).

_FEEDBACK