Trying to pick a side in the open-source versus closed-source AI fight? Here’s the fast way to think about it.
I kept seeing this debate flare up and it felt too tangled to follow. Then I watched a breakdown from the creator of the Forward Future newsletter, who walks through both camps and lands on a clear recommendation. I think it’s the cleanest explainer I’ve seen, so let me hand you the short version.
First, the setup. Nvidia’s CEO published a pro-open-source letter, and nearly every major tech CEO co-signed it. One company stayed quiet and went the other way: Anthropic, arguing open models are too dangerous. That’s the whole tension in a nutshell.
Quick definition from the author: open source means the model is public to run, inspect, and modify. Closed source means you go to one company for intelligence, and if they change terms, you’re cut off. Same tech under the hood. The only real difference is the business model.
🟢 The case for open source (the author’s read)
- More competition, so lower prices for us
- More eyeballs on the code, which historically means better security
- Transparency instead of “just trust us”
- Less concentration of power in two companies
- Local use means real privacy
🔴 The case against (Anthropic’s position)
- Anyone can strip out the safeguards
- Once weights are released, you can’t pull them back
- It may lower the barrier for bad actors
- Responsibility gets murky when something goes wrong
Here’s where the expert pushes back. Closed models get jailbroken too. Dangerous info already lives on the open internet. And running a truly powerful model takes data-center-scale compute, which means a big provider is in the loop and accountable, not some hacker on a laptop. His point: the safety gap is smaller than the marketing suggests.
The part that reframed it for me: who actually wins if open source takes off. The creator maps the AI stack and argues chips, energy, data centers, apps, and tooling all benefit from cheaper, more-used intelligence. That ties into Jevons paradox: as intelligence gets cheaper, total usage skyrockets. The one layer that feels the squeeze is pure model providers selling tokens at fat margins. His contrarian take is that even OpenAI and Anthropic will be fine, because their apps, memory, and user experience are far ahead of anything you’d self-host.
Then there’s China. The author explains they lead on open models because they have the researchers and the energy but not the top chips, so giving models away is a classic play to pressure the leaders’ margins. He also demystifies “distillation”: one model teaching a smaller one. It’s common and legal in general, though it can cross into IP theft if it violates a company’s terms of service.
His recommendation
- Don’t ban Chinese open models. They pressure US prices and spread access, which he cares about most.
- Fund US open-source AI like the basic research it is.
- Require safety testing, but keep it cheap enough that startups can actually compete. Otherwise it’s regulatory capture.
One implementation tip he offers for the distillation problem: providers should use KYC (know your customer), the same approach banks use, instead of pushing for bans.
The video wraps with Anthropic’s CEO responding that they never called for a ban, just mandatory safety testing. The author’s honest verdict is that he still disagrees, since he thinks broad access helps defenders more than attackers.
My takeaway? If you’re choosing where to build, closed models still win on polish today, but open models are close and getting cheaper fast. Lean open for cost, control, and privacy; lean closed for the smoothest experience.
Watch the full video for the stack breakdown and the Anthropic letter walkthrough. It’s worth your time.