AI Is Hitting the Gas and Brake

Welcome back apprentices! 👋

Hey {{first_name|friend}},

AI had a very normal week: 

slam the brakes, hit the turbo, pick a fight with GitHub, and casually wander into a biology lab.

What looks like a handful of unrelated headlines is actually pointing to something bigger.

_{Read Time: 5 minutes}

🪄 Janie Wasn’t Real. A 19-year-old redhead named Janie pulled nearly 1M TikTok views in a week with sorority recruitment videos — except Janie was entirely AI, created by a16z partner Olivia Moore for about $100. Moore spent roughly 30 minutes a day animating one ChatGPT-generated image with MiniMax and Grok Imagine, and even after viewers spotted the AI clues — and TikTok labeled 8 of 20 videos — the views kept coming. When Moore finally revealed the experiment, she said most reactions were surprisingly positive, raising an awkward new question for the creator economy: does the person behind the content actually need to exist?

🎬 Hollywood Draws the AI Line. ByteDance has agreed to new copyright guardrails for its Seedance and Seedream AI models after a viral Tom Cruise deepfake triggered Hollywood’s first major cease-and-desist against an AI company. The protections will also apply across apps like TikTok, CapCut, and Dreamina — a sign that AI video has officially graduated from “weird internet toy” to “call the lawyers.” But with rivals like Kling and Alibaba’s Wan still advancing fast, Hollywood’s copyright fight is far from over.

🧠 Doctor Cracks a 20-Year Math Problem. Beijing neurosurgery resident Shanmu Jin used GPT-5.6 Sol to produce a proof for Crouzeix’s Conjecture, a matrix problem unsolved since 2004, after a 16-hour autonomous run. Jin isn’t a professional mathematician, yet leading experts — including the conjecture’s creator — have checked the proof, with formal peer review still pending. The twist: AI may be making advanced discovery more accessible, while turning human verification into the increasingly valuable part of the equation.

Find out if another carrier could offer you a better rate. Compare your offers all in one place and get your best rate on LendingTree. Just answer a few questions and get your quotes in a matter of minutes.

Check My Quotes

AI’s frontier is doing something strange: getting dramatically faster while its creators simultaneously install bigger emergency brakes. 

OpenAI temporarily paused reinforcement-learning training and is holding back its largest planned frontier run over cybersecurity and alignment concerns — while separately previewing an API tier that makes GPT-5.6 Sol up to 14x faster. Meanwhile, Cursor is moving beyond writing code into hosting it, putting its new Origin platform on GitHub’s turf, and Anthropic is making a different bet: that AI will earn public trust through useful scientific results rather than better PR. Its latest protein-design experiments offer an early example, with Claude coordinating complex research workflows that produced experimentally validated molecules.

Put it together and the AI race is becoming a race across speed, safety, infrastructure, and real-world usefulness.

If this week’s OpenAI news sounds contradictory, that’s because it is — and that tension may define the next phase of AI.

On one hand, OpenAI is slowing frontier development. After a security incident and early signs that Astra could reach top cybersecurity capability levels, the company paused reinforcement learning for two weeks, tightened research environments, expanded monitoring, and held back its largest planned RL run. Its systems can now escalate suspicious behavior to automated investigators and humans, with critical alerts potentially triggering a pause within 30 minutes. Monitoring alone can add about 20% extra compute cost.

That’s not minor.

On the other hand, speed is becoming a product.

Ultrafast runs GPT-5.6 Sol at up to 750 tokens per second, aimed at use cases where latency is the bottleneck: security response, finance, voice, coding, and rapid iteration. Internally, it has already compressed workflows that used to take much longer.

The key takeaway: faster AI and slower AI development are no longer opposites.

Inference is accelerating while training is becoming more cautious. As models grow more capable, the safety infrastructure around them becomes heavier.

The frontier is shifting from build → test → launch to build → monitor → monitor the monitor → test → maybe adjust → launch.

Progress, but with more overhead.

AI has already changed how software is written. Cursor now wants to change where it lives.

Its Origin platform combines repos, pull requests, browsing, GitHub sync, and AI agents in one system. GitHub repos can be mirrored so developers don’t have to fully switch immediately.

That matters because GitHub’s real advantage is inertia — the world is already there.

So Cursor doesn’t force a migration. It brings GitHub along.

The bigger shift is what happens when repos are built for AI agents, not just humans. Instead of passive code storage, they become active environments where AI reviews, edits, tests, and eventually helps maintain software continuously.

Competing with GitHub then isn’t about storage.

It’s about owning the workspace where humans and AI build software together.

Dario Amodei’s argument is simple: trust AI less in words, more in outcomes.

His example — curing cancer — is intentionally extreme, but the point is that public trust will depend on real-world results, not promises.

Anthropic’s latest research moves in that direction.

Claude was given tools, internet access, and computational resources to run parts of a protein-design workflow. It produced candidates that worked for 14 of 15 targets, with hit rates of 22–35% depending on setup.

Important caveat: Claude didn’t run a lab. Humans still synthesized and tested everything, and specialized tools did much of the heavy lifting. But the key result is that a general AI system coordinated a complex scientific pipeline with limited human input.

That matters because it suggests a shift in AI’s scientific role.

Not a single “Eureka” moment, but thousands of small decisions: selecting tools, running analyses, iterating failures, organizing results, and freeing researchers to focus on which problems are worth solving in the first place.

And that may be a bigger change than any demo.

Even if you never train an AI model, host a repository, or design a protein, these developments are pointing toward what your everyday AI tools could become.

First, waiting may disappear. If frontier-level models can respond close to real time, AI agents become far more practical inside phone calls, customer service, research, coding, and other workflows where a 30-second thinking break currently feels like an eternity.

Second, AI products are starting to absorb entire workflows. Cursor moving from coding assistant to code hosting is a good example. The competitive advantage may increasingly belong to companies that own not just the AI model, but the place where the work itself happens.

Third, safety may start affecting release schedules. OpenAI’s decision to hold back parts of frontier training shows that increasingly capable systems can create costs and delays that benchmark charts don't capture. The fastest lab on paper may not always be the first one comfortable shipping.

And finally, AI’s biggest test is shifting from “Can it answer this?” to “Can it accomplish something useful in the real world?”

Writing a clever email is nice.

Helping scientists run experiments, engineers resolve outages, or developers maintain entire codebases is where the economic — and societal — stakes get considerably larger.

When the next AI announcement claims something revolutionary, ask three boring questions: What did the AI actually do? What did humans still have to do? And did it work outside the demo?

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 👀

Practice Docker, Kubernetes, Linux, Terraform, and Git in real terminal environments. 

Explore hundreds of hands-on labs across the entire DevOps stack that let you provision clusters, automate infrastructure, and manage deployments. Experiment freely across multiple playgrounds that offer sandbox environments. 

Learn DevOps by doing, not just videos or theory.

Start For Free

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 Jake for this creation!🥳

Want to be featured next? Keep those generations coming!

Inside a vibrant photography studio, [your object] is positioned at the center of a rotating circular stage. Around it, six professional cameras capture the same object simultaneously, each from a completely different perspective: extreme close-up, overhead, low-angle, macro, profile, and wide environmental shot. Large prints from each camera surround the studio, creating a visual mosaic of the same object seen in radically different ways. Bright colored gels cast cyan, violet, pink, and amber light across the scene, while the central object remains perfectly sharp. Ultra-detailed, photorealistic, cinematic studio photography, crisp textures, rich reflections, high-end editorial aesthetic.

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