Open AI vs Closed AI: Which Actually Wins in 2026? | Ep 4

00:00:00:03 - 00:00:01:19

Two years ago, if you

00:00:01:19 - 00:00:02:01

wanted

00:00:02:05 - 00:00:06:22

the best AI on the planet, you paid either anthropic or open AI.

00:00:06:22 - 00:00:09:00

That was the conversation.

00:00:09:00 - 00:00:14:20

Open source. Was the bargain bin fine for tinkering? Never for production.

00:00:15:01 - 00:00:25:23

Today, a free model you can download tonight beats the best paid model from a frontier lab on real world software engineering benchmarks

00:00:26:00 - 00:00:27:05

the gap on

00:00:27:07 - 00:00:31:19

general intelligence benchmarks went from 17.5 percentage points

00:00:31:20 - 00:00:39:13

to about 3/10 of 1% in a single year. The moat didn't erode. It collapsed.

00:00:39:15 - 00:00:43:21

here's what nobody on either side of this debate wants to say out loud.

00:00:43:21 - 00:00:46:08

I want to welcome you back to super confident AI

00:00:46:11 - 00:00:50:08

today. The real state of open versus closed war.

00:00:50:10 - 00:01:03:12

Let me start with the number that should be on every founders whiteboard. ChatGPT 5.2 runs about $1.75 per million input tokens, $14 per million output tokens

00:01:03:15 - 00:01:09:03

Run a meaningful production workload through that. You're looking at around $2,300 a month.

00:01:09:03 - 00:01:12:10

Same workload on Deep Seek v 3.2

00:01:12:13 - 00:01:16:22

through deep infra, you're about $168 a month.

00:01:16:23 - 00:01:18:23

That's 13 times the cost

00:01:19:00 - 00:01:23:23

13 for workloads where honestly, your users can't tell the difference.

00:01:24:01 - 00:01:25:20

I run a model agnostic platform.

00:01:26:00 - 00:01:28:03

We root across more than 100 models.

00:01:28:05 - 00:01:29:12

I'll tell you straight,

00:01:29:14 - 00:01:42:12

most builders running closed only stacks or not paying for quality. They're paying for the feeling of safety. And in a market this competitive, that feeling is the most expensive luxury you can buy.

00:01:42:14 - 00:01:54:10

Sam Altman himself said OpenAI might be, in his words, on the wrong side of history. When the CEO of the most famous closed lab in the world says that you should at least raise an eyebrow.

00:01:54:12 - 00:02:06:09

I'm not here to be a homer for open source. Let's be honest about what close still does better. Open SWE bench verified. That's the benchmark for fixing real GitHub issues.

00:02:06:13 - 00:02:14:13

Claude 4.5 sits at 77.2%, GPT 5.1 at 76.3.

00:02:14:14 - 00:02:26:04

The best open models Lama for Deep Seek V4 are at 72. That's a real gap if your product lives at the frontier of a joint.

00:02:26:06 - 00:02:35:18

closed also leads in polished multimodal, in safety tuning, in the kind of integration where you don't want to manage inference infrastructure yourself.

00:02:35:18 - 00:02:39:12

If you absolutely need the bleeding edge, just pay the tax

00:02:39:12 - 00:02:45:02

for the other 80% of products. You're overpaying to feel comfortable.

00:02:45:07 - 00:02:48:18

Okay, here's where I lose half the open source crowd.

00:02:48:21 - 00:02:51:04

Open weights have a real problem.

00:02:51:04 - 00:02:54:19

It's called malware, and we need to stop pretending it isn't real.

00:02:54:19 - 00:03:02:11

Security researchers found that deep seek out of the box will actually help users generate malware and viruses.

00:03:02:14 - 00:03:15:19

Better aligned closed models like OpenAI's oh three mini have to actively be jailbroken to do the same thing. That's not a small difference. That's basically the difference between a lock storefront and an open warehouse.

00:03:15:21 - 00:03:19:03

When you release open weights, you release them forever.

00:03:19:05 - 00:03:24:07

Anyone with a few GPUs can fine tune the safety training right out of the model.

00:03:24:09 - 00:03:30:09

That's not a bug of open source. That's the cost of openness, and we should be honest about

00:03:30:11 - 00:03:35:09

Second piece the repositories themselves are now an attack surface.

00:03:35:10 - 00:03:38:01

I 2026 X-Force

00:03:38:01 - 00:03:42:17

report flagged nearly four times increase in supply chain compromises.

00:03:42:17 - 00:03:43:23

Since 2020,

00:03:44:01 - 00:03:49:02

open source model repos have become a prime vector for malware delivery.

00:03:49:04 - 00:03:55:06

Attackers post poison model files that execute arbitrary code the moment you load

00:03:55:07 - 00:04:05:11

The 2026 Open Source Security and Risk Analysis report, the Ausra found vulnerabilities doubled to 581 per code base.

00:04:05:15 - 00:04:16:10

87% of code bases are at risk. 17% of open source components are now entering code bases outside standard package managers

00:04:16:13 - 00:04:24:16

copy pasted snippets, AI generated code, direct vendor inclusions. They're all invisible to your traditional scanners.

00:04:24:16 - 00:04:30:10

Third piece AI generated malware is breaking defenses we've relied on for decades.

00:04:30:12 - 00:04:40:08

Signature based antivirus works by recognizing known patterns. AI generated malware can be structurally different every single time it generates.

00:04:40:09 - 00:04:45:21

The cat and mouse game just turned into a cat and a shapeshifter

00:04:45:23 - 00:04:57:16

Look, I'm not anti. My business runs on open routing, but anyone selling you open source AI as a pure democratization story is only selling you half the truth.

00:04:57:21 - 00:05:02:19

The honest version. Open source AI works exactly like open source software.

00:05:02:20 - 00:05:05:03

Powerful democratizing? Yes,

00:05:05:03 - 00:05:09:20

and it shifts a real share of the security burden onto whoever deploys

00:05:09:20 - 00:05:12:16

If you don't have the team to vet the weights,

00:05:12:20 - 00:05:20:06

you sandbox inference and monitor outputs. You don't have an open source strategy. You have a liability.

00:05:20:08 - 00:05:21:23

The question I'll leave hanging?

00:05:21:23 - 00:05:37:11

Should there be a minimum responsible disclosure standard for open model releases like the security industry has for vulnerabilities? Or does any gatekeeping just become moat building by the closed labs dressed up in safety theater?

00:05:37:15 - 00:05:40:19

Drop your answer in the comments. I'm genuinely curious.

00:05:41:00 - 00:05:43:15

Here's a double shot on two related issues.

00:05:43:18 - 00:05:45:23

The first geopolitics.

00:05:46:03 - 00:05:48:18

The biggest open weight releases right now.

00:05:48:20 - 00:05:53:23

Deep seek. Quinn, Kimi, GM they're all Chinese.

00:05:54:00 - 00:05:57:06

The biggest closed source players, they're all American.

00:05:57:08 - 00:06:05:17

Open source AI became part of US China geopolitical competition after Deep Seek released R1 in January

00:06:05:19 - 00:06:11:13

Is open source. A Trojan horse was the only thing keeping American closed labs

00:06:11:15 - 00:06:14:08

Both can be true and they likely are

00:06:14:10 - 00:06:19:20

two open washing read the license, not the press release.

00:06:20:00 - 00:06:28:03

Meta calls mama open source, but the license actually restricts commercial usage over 700 million monthly users.

00:06:28:05 - 00:06:32:00

OpenAI flirts with weights, never the training data.

00:06:32:03 - 00:06:40:22

There are three definitions of open weights, code and data. And most companies ship one and call it a revolution.

00:06:41:01 - 00:06:42:10

If you're a founder,

00:06:42:12 - 00:06:44:17

ask three questions before you

00:06:44:19 - 00:06:50:16

Can I see the weights? Can I see the training data? Can the provider yank access from me tomorrow?

00:06:50:18 - 00:06:53:02

You'll know pretty fast where you actually stand.

00:06:53:04 - 00:06:56:17

Here's how I think about it on my platform. Three lanes,

00:06:56:19 - 00:07:07:18

Lane one frontier agent work that cannot fail. Legal review. High stakes coding close models every time. Just pay the tax

00:07:07:20 - 00:07:13:05

lane two high volume consumer workloads where margin is the game.

00:07:13:08 - 00:07:18:01

Open weights with a closed model fallback. This is where you actually build your business.

00:07:18:03 - 00:07:29:00

Lane number three sensitive enterprise data financial, medical, all regulated open weights on your own infrastructure with real supply chain

00:07:29:02 - 00:07:33:02

sign model cards, vetted weights, sandbox the inference

00:07:33:06 - 00:07:36:05

you're going to want to route by the job, not by the

00:07:36:06 - 00:07:38:22

The next 12 months are not open versus closed.

00:07:38:22 - 00:07:45:11

They're smart routers versus dumb stacks. Founders who build optionality into their architecture

00:07:45:11 - 00:07:48:15

will outlast founders who picked a team and prayed.

00:07:48:17 - 00:07:51:11

if this episode hit. Smash the like.

00:07:51:12 - 00:07:52:10

Drop a comment

00:07:52:10 - 00:07:56:02

with which lane your product lives in and I'll see you on the next

00:07:56:03 - 00:07:58:03

This is super confident AI.