In partnership with
My fellow AI explorers
Bill Gates just told the world AI needs guardrails, in the same week OpenAI dropped a chip benchmark timed almost perfectly to rattle Nvidia's earnings call. Two very different power players, one very same message: nobody who actually builds this stuff wants you to relax.
In today’s edition:
🏛️ Bill Gates says AI needs real limits, including jobs humans keep forever
🌶️ OpenAI's Jalapeño chip just beat Nvidia on its own turf, hours before earnings
🎙️ Google's new transcription model, Nvidia's earnings, and Meta's AI layoff plan that fell apart
Your Competitors Just Hired an Entire GTM Team. No Humans.
AI agents are doing the work of full GTM teams. Pipeline, content, customer service. The startups using them aren't waiting for headcount approvals.
Get the free Practical Guide to Agentic GTM for Startups and see exactly where to start.
30-Second AI Play
🎬 Turn Claude Into a Free Motion Graphics Studio (No After Effects Needed)
Someone just built a fully color-graded, animated video with money counters, motion backgrounds, and a map animation using only Claude and CapCut. Zero After Effects. Here's how to steal the workflow:
Grade your footage first. Ask Claude to generate a custom LUT file with a prompt like "clean soft pastel LUT, cinematic and clean." Download it, import it into CapCut under Adjustment → LUT, and dial the intensity to taste.
Animate your numbers. Ask Claude to generate an SRT file counting from 0 to whatever number you need, drop it straight onto your CapCut timeline, then group the clips into a compound clip and speed-ramp it for a slick counting effect.
Install Claude Code, Node.js, and Remotion for anything more advanced, like motion backgrounds, animated graphics, or map animations. This is the free-version ceiling, so you'll want the paid Claude plan for this part.
Feed it a reference image. Grab a Pinterest screenshot of the style you want, drag it into Claude Code, and describe exactly what you want built, aspect ratio, duration, and all.
Export with a transparent background so you can layer the finished graphic directly over your existing footage in CapCut.
🔍 Pro tip: end every prompt to Claude Code with "ask me any clarifying questions so we nail this spot on." It forces Claude to interview you about stacking order, timing, and motion style before it spends four minutes generating the wrong thing.
Want more workflows like this one broken down step by step?
Then The Operator Brief is for you! We are currently running a welcome offer to the first 250 people that join!
THE BIG STORY
Bill Gates Wants Some Jobs Legally Off-Limits to AI
Bill Gates just published a 6,000-word essay arguing that AI needs "significant limits," or the harm to humans will outweigh the good.
Here's what he's actually proposing:
A category he calls "Human Reserved" jobs: work that society deliberately keeps for people, even when AI or robots could technically do it better. He compares it to a nature reserve, land that could be developed but is protected on purpose.
Childcare and jury service are two examples he's floated as clear candidates. In education and healthcare, he says you wouldn't ban AI outright, but the profession itself could stay human-led while workers use AI to boost their output.
The protection could be permanent ("this job is done by humans, forever") or temporary, buying time for workers who realistically can't pivot into a new career, like manufacturing workers late in their careers.
He also wants governments to tax AI tokens and robots the way payroll tax works on a human employee, to remove the financial incentive companies currently have to automate everything.
In a separate GeekWire interview, Gates tore into the U.S. government's current voluntary pre-release testing framework, calling it toothless: "What is the threshold that's being examined, and what is the action taken when you cross that threshold? The whole thing seems so empty to me."
He's reportedly seeking a meeting with Xi Jinping to push for coordinated global AI safety rules, betting China might actually agree to restrict certain dangerous model releases.
Here's the uncomfortable part. This is the same Bill Gates who, back in February, told Jimmy Fallon that AI would replace doctors, teachers, and most professionals within a decade, and that humans "won't be needed for most things." Now he's the one arguing for legally protected human-only work. That's not a small pivot. It's one of tech's most influential voices going from sounding the alarm bells to admitting the default path he helped champion needs a fence around it, and to starting to do something about it.
*Not everyone will read "Human Reserved" as protection, though. The moment you write a list of jobs AI legally can't touch, you've quietly signed off on AI taking everything not on that list. And somebody has to write the list, which means somebody gets to decide whose livelihood counts as worth saving and whose doesn't. That's a political fight dressed up as a labor policy.
🔮 Prediction: Expect the "Human Reserved" idea to move faster through legislatures than the AI tax ever will, and here's the reasoning. A token tax on AI usage requires international coordination, because any company facing it can just shift workloads to a jurisdiction that doesn't tax it, the same problem that's stalled corporate tax harmonization for decades. A reserved-jobs law doesn't have that problem. A single country, state, or even city council can pass one unilaterally through ordinary labor law, no treaty required, no tech company has to change a single line of code, and politicians get a visible, easy-to-explain win heading into an election. That combination, low coordination cost plus high political payoff, is exactly the recipe that gets bills moving.
Watch for the first attempts to land in childcare licensing, eldercare staffing ratios, and courtroom or jury-adjacent roles, the same places Gates named. Expect the fights that follow to look a lot like the gig-worker classification battles of the 2020s: constant argument over what actually counts as "the reserved job" versus "AI-assisted work that happens to be done by a human," with lobbyists on both sides fighting over the fine print long after the headline law passes.
Would you want your own job legally protected from AI, even if AI could do it better or cheaper? Hit reply and tell me where you land.
THE BIG STORY
OpenAI's Jalapeño Chip Just Beat Nvidia, One Day Before Earnings
OpenAI unveiled its first custom AI chip this week, and the timing was not an accident.
Here's what you need to know:
The chip is called Jalapeño, co-developed with Broadcom on silicon and networking and Celestica on systems integration, part of an October 2025 deal to build 10 gigawatts of custom AI accelerators together.
OpenAI's own benchmarks show Jalapeño delivering 1.5x to 1.9x more throughput at peak load versus Nvidia's Blackwell-class systems, with 1.7x to 3.6x lower latency, tested across GPT-OSS-120B, DeepSeek R1, and Kimi K2.5.
It's built for inference only, not training. Each rack packs 128 accelerators, 1.7 exaFLOPS of 4-bit compute, and 27.5 TB of HBM4 memory, running at a comparatively lean 700 watts.
Analysts told CNBC this is a genuine "threat to Nvidia's inference margins," specifically, the fastest-growing corner of the AI compute market, even though Nvidia still controls the vast majority of total AI compute and holds a deep moat through CUDA.
The timing lines up almost too well: OpenAI dropped this one day before Nvidia's own Q2 FY2027 earnings, reporting after market close today, with Wall Street expecting roughly $92 billion in revenue and laser focus on Q3 guidance.
OpenAI isn't going all-in on its own silicon either. Hardware chief Richard Ho said plainly: "We're going to need a lot of compute. Jalapeño is part of that. Nvidia is part of that. AMD is part of that." Gen 2 and Gen 3 chips are already in development.
Here's the uncomfortable part. This wasn't a quiet engineering blog post. It landed the day before the single earnings call the entire AI trade is built around. And it targets the exact metric, inference margin, that analysts are about to grill Jensen Huang on tonight.
Worth flagging: Jalapeño was benchmarked against Nvidia's Blackwell, not Nvidia's newer Vera Rubin platform, which only recently started shipping. That makes the comparison flatter OpenAI's timeline more than it reflects a true apples-to-apples fight against Nvidia's current lineup.
🔮 Prediction: Nvidia's numbers tonight will almost certainly beat estimates. That's not in question, since Jalapeño doesn't actually deploy until the end of the year and hasn't touched a single dollar of this quarter's revenue. The real signal will be in the Q3 and Q4 guidance language and whatever Jensen Huang says on the analyst call about custom silicon, because if Nvidia feels the need to address Jalapeño unprompted, that confirms it landed.
Here's the reasoning chain: inference is growing faster than training as more of AI spend shifts from building models to actually running them for users, which means whoever controls inference economics increasingly controls where the next round of hyperscaler capex flows.
Expect Nvidia's response to lean hard into Vera Rubin's raw performance numbers and CUDA lock-in rather than a direct benchmark rebuttal, because a spec war against a chip built specifically for OpenAI's own workloads is a fight Nvidia doesn't need to pick in public. Expect other hyperscalers to fast-track their own custom silicon announcements to ride this same news cycle. Anthropic already confirmed it's exploring an in-house chip. By sometime in 2027, expect buyers to increasingly split workloads, training on Nvidia, more inference on custom silicon, which won't shrink Nvidia's revenue, but will visibly compress its growth rate in the metric investors actually watch.
Do you think custom silicon like Jalapeño actually dents Nvidia's dominance, or is this a benchmark that ages badly the moment Vera Rubin ships at scale? Reply and tell me where you land.
Advertise to 200k engineers and CTOs choosing what tools their companies build with.
Other Relevant AI News!
🎙️ Google just launched Gemini 3.5 Transcribe, a new speech-to-text model that cleans up filler words, handles self-corrections mid-sentence, and identifies up to three speakers with word-level timestamps. Check it out here.
📊 Nvidia reports Q2 earnings tonight after market close, with Wall Street expecting around $92 billion in revenue and every analyst question likely circling back to custom silicon competition.
🤖 A Reuters investigation reveals Mark Zuckerberg quietly planned to shrink some Meta teams by up to 60% under a project code-named "Project OT," before scrapping the second wave hours before the first layoffs hit, after employee revolt and AI agents that didn't deliver.
Golden Nuggets
🏛️ Watch which government moves first on a "Human Reserved" jobs law. That's the real signal, not Gates' essay itself.
🌶️ If Jensen Huang brings up custom silicon on tonight's call without being asked, that's your confirmation Jalapeño actually landed.
🎙️ Voice is quietly becoming the default AI interface. Transcription models are the plumbing nobody notices until it breaks.
Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)
Until our next AI rendezvous,
Anthony | Founder of Uncover AI