Ocean Bennett (0CE4N ON SPOTIFY)

Two months ago, I launched icebergcreator.com Early Success Metrics: • Total Impressions: 3,030 • Total Clicks: 404 • Average CTR: 13.3% This performance places the site in the top 1-5% of new websites, driven by a high-value, U.S.-centric audience. #SEO #SaaS #FREE

@M1ndPrison Agents only became practically useful ~6 months ago. We’re still on a steep exponential, so most of what feels like a hard limit today will probably look temporary soon. Computer science has always been about continuously learning, and unlearning assumptions.

"Once that’s done, the system runs itself." What does this system do? Does it build new features by itself? Does it explore new concepts? Test new ideas? Anything fully automated isn't going to have any of those things. It cannot do anything new. It can only do what was concretely designed into goals, harnesses etc. We can only automate what humans have already concretely defined. That significantly limits the scope of full automation.

That's impossible. No matter how far you scale pattern-matching machines, you cannot escape Polanyi's Paradox. All of the things of value are locked within intelligent minds of which AI lacks. You cannot automate this facet. Which means everything of value will still require as much input from humans as it ever did.

@nulonaw @ThePrimeagen Not about tool usage. The edge is in designing the systems themselves, the orchestration, evals, state management, and feedback loops that make prompting almost unnecessary. Tools will spread; the ability to build and iterate those environments won’t equalize as fast.

@oceanbennett @ThePrimeagen Nope. Tool usage won't give you any advantage. Whatever works well will be propagated to others and will be easily accessible. That also has been the trend in the last 2 years.

Something that has greatly improved the reliability of the code I produce via LLM is a technique I took from cursor. Field guide. First, how I start. I nitpick everything and I always make it produce the exact code I want, I use voice dictation and talk for quite a bit. Second, every single back and forth is recorded to the markdown file. Along with each commit hash for the change. Then I have the agent review the back and forth and the changes and come up with durable and generalized feedback that it stores in a directory (field guide). The directory has init.md that is just a bunch of links and short descriptions for each file. This way I don't blow the context, but instead it can do some basic tool calls and operations to find exactly what it's looking for. Over the last 10 sessions, the in-depth feedback I've had to give has greatly reduced despite no change in model or harness that I'm aware of. Now the feedback is project specific, and is not universal. I am okay with that.

@M1ndPrison You’re still treating it like a chat that needs constant babysitting. Codify the purpose into versioned docs and wire them into the CI/CD pipeline. Once that’s done, the system runs itself. You’re maintaining the system, not prompting it. That’s the shift.

You don't do this once and stop. Those goals can't be abstract-self-expanding tasks. It will always be narrow in nature. If you are interacting with an AI system, the input is the prompt. That's the architecture no matter what layers it is wrapped in. All new semantic meaning has to enter the input for any valuable output. The only source of that remains human beings.

No. Setting up the autonomous system is not “deciding what goes in the prompt.” Once the harness, tools, goals, and loops are built, the system runs with near-zero ongoing prompts. Calling every code harness “prompt management” is just redefining the entire architecture as the one thing you want it to still be. The leverage is that the human stops managing prompts and starts managing systems that no longer need them.

"The input just shifts to higher-leverage work: setting up autonomous systems, defining goals, and applying judgment." And pretty much all of that comes down to deciding what goes in the prompt and how it gets there. All of the code harnesses are essentially prompt management frameworks.

@M1ndPrison Not full circle lol. Engineers stay in the loop, that’s the point. The input just shifts to higher-leverage work: setting up autonomous systems, defining goals, and applying judgment. Automation creates that space. Same loop, different center of gravity.

@oceanbennett Sure, and now we have come full circle. Which is engineers will still be required to supply substantial input to these systems for any high value output.

Humans are embodied, goal-driven prediction engines with biological conditioning and subjective awareness. We do predict words and actions based on past patterns, but those predictions are driven by biology, emotion, and a desire to navigate a real physical world

@DaveBlundin Still fully on board that America needs more engineers. Worth noting China’s graduate glut has left many STEM/PhDs underemployed too, scale helps, but matching talent to real jobs is the harder part.

China is now educating 10x more engineers than the US, double the PhD STEM students, and holds the #1 spot globally in patents. Countries are now sending their best minds to China to learn. That's a very different story than the media projects! My full episode with Joseph Aoun, President of @Northeastern, just dropped and it’s chalk full of brilliant nuggets.

@M1ndPrison Human engineers stay in the loop, that’s exactly why AI will handle high-value novel work. Automation clears the routine so we focus on the highest-level judgment and creativity.

@oceanbennett They learn patterns of text, and at massive scale, that becomes very useful. Nonetheless, none of that ability comes from intelligence. They will continue to improve, but full self-automation won't happen for high value novel work.

Such a bad take. Belief that open source should exist doesn’t equate to belief that literally everything should be open source. And NVIDIA invested $20b into open source. Embarrassing.

@elonmusk @BillyM2k Only if LLM outputs are fully citeable and verifiable, like Lean proofs.

@BillyM2k Your future doctor will be AI, but probably not of any name people know today

@MikeWithAHotDog @ThePrimeagen Devin already handled Nubank’s multi-million-line production ETL migration. Keep building the hard things the slow way.

@oceanbennett @ThePrimeagen I can't hear you over actually building hard things

@MikeWithAHotDog @ThePrimeagen legacy cope

@oceanbennett @ThePrimeagen That's hilariously disconnected from reality

@MatthewBerman To me, opposing open source means opposing human progress. Democratizing access allows more minds to contribute, accelerating innovation for everyone. Even big tech relies on open source to advance, and always has.

Banning Chinese open source seems worse than bolstering our own open source efforts. Open source AI is basic research.

@Pirat_Nation The 300M+ budget model is a temporary peak. As the cost of intelligence drops and agentic workflows streamline production pipelines, game development costs should plummet. The real question is whether studios pass those savings on or just inflate game scope even further.

David Jaffe, the creator of God of War, says the move away from physical games isn’t driven by greed but by the economics of modern game development. According to Jaffe: “Here’s the actual uncomfortable truth: Phasing out physical media is not some evil, greedy, illuminati anti-consumer plot. It is just the brutal unavoidable financial reality of an industry trying to survive…when games cost $300 M+.” He also argued that third-party publishers aren’t pushing back against Sony’s shift toward digital because it benefits them: “You can see why third-party publishers are not fighting Sony on this, they’re popping champagne.” Jaffe believes that with AAA game budgets now regularly exceeding $300 million, publishers are looking for ways to cut costs and maximize revenue, and reducing reliance on physical media is one of those strategies.

When systems thinkers leverage technology, the entire game changes. Linear thinkers and traditional 'book-smart' approaches were built for a legacy world, not the future.

@thsottiaux @betomoedano I got you Tibo: Only if you ignore GPT-5.6 Luna ($1/$6) and Terra ($2.50/$15). The chart only showed Sol.

@betomoedano Who is gonna tell him

Opus 5 is now the cheapest frontier model

@Polymarket The real benchmark will be seeing if this level of mathematical capability can produce novel, publication-grade CS algorithms.

JUST IN: Anthropic reveals Claude Opus 5 achieved a perfect 42/42 score on the 2026 International Mathematical Olympiad.

@unclebobmartin Engineering has always been about solving problems and optimizing systems, not grinding out syntax at a keyboard.

I'm starting a brand new project from scratch. It's just so much fun to watch the agents work out all the wiring issues. This saves me SO MANY headaches.