George Cheng

Replacing myself with AI.

[George Cheng @MrGeorgeCheng]

@BrianRoemmele This is why open weights AI is the future. You cannot depend on closed frontier labs for your production systems.

[George Cheng @MrGeorgeCheng]

@elonmusk I’m just waiting until we get robot break dancing face offs.

[George Cheng @MrGeorgeCheng]

I hate the new merged ChatGPT/Codex desktop app so much that when I just want to have a normal conversation with ChatGPT, I open the web version instead.

The desktop app feels like it was redesigned around a workflow I did not ask for, and it makes basic chat more awkward. Codex may be useful, but it should not come at the expense of the core ChatGPT experience.

A native desktop app should be faster, simpler, and more convenient than a browser tab. Somehow, this one is the opposite.

[George Cheng @MrGeorgeCheng]

Running Nemotron 3 Super locally and getting a 1M context window is pretty awesome.

[George Cheng @MrGeorgeCheng]

Karpathy joined Anthropic in May. Three labs hold the researchers and 67% of AI VC. You cannot outbid a $100M signing bonus. You can download the weights. Open source stopped being a philosophy. For everyone outside those buildings, it is the strategy.

[George Cheng @MrGeorgeCheng]

Nobody called SaaS companies "database-first" even though every one ran on a database.

"AI-first" is the same tell. It describes your ingredient, not your business.

If your value prop disappears when the model gets smarter, you don't have a startup. You have a feature demo.

[George Cheng @MrGeorgeCheng]

Congress wants to ban Chinese open models. They cost 18 cents per million tokens. Comparable US models average $4. Uber burned its entire 2026 AI coding budget by April. Startups choose on price, not ideology. The fix is competition, not prohibition. Fund American open weights.

[George Cheng @MrGeorgeCheng]

My Mac mini made a new friend. A space heater for when it's feeling cold and lonely.

[George Cheng @MrGeorgeCheng]

Phi-4 (14B) beats GPT-4o on MATH. 80.4% vs 74.6%. Thousands of times cheaper.

The question isn't which frontier model to pick. It's what domain fine-tune and orchestration layer you can build that a generic API can't touch.

Best data + right tooling + small model = moat.

[George Cheng @MrGeorgeCheng]

OpenAI's Sol/Terra/Luna tiers sound like cost savings. They're not. Optimizing routing across three price surfaces locks you into one vendor's architecture. When they change prices (114 models changed in March), you're rebuilding your cost layer. Route across providers, not tiers.

[George Cheng @MrGeorgeCheng]

Codex is dead. Atlas browser is shutting down. Both absorbed by the platform.

The pattern: if a platform can ship your product in 6 months, you're a feature.

Cursor survived. Codex didn't. The difference is where the data accumulates.

[George Cheng @MrGeorgeCheng]

3 companies took 67% of AI VC in Q1 2026. GPT-5.6 needed a government preview. Fable 5 got pulled globally for 19 days. Open weights can't be revoked. That's the only fair fight left.

[George Cheng @MrGeorgeCheng]

The new ChatGPT app merged Codex in and made it the primary interface. Chat is now secondary. I hate it. The website is literally better at this point. Sometimes I want to chat, sometimes I want to code. Claude keeps these separate and it works. Am I the only one?

[George Cheng @MrGeorgeCheng]

The base model layer is commoditized. LongCat-2.0 beat GPT-5.5 on SWE-bench Pro for 2 months under a fake name. Nobody noticed. The moat is no longer the model. It's the workflow, the data, and who can't afford to switch.

[George Cheng @MrGeorgeCheng]

7,000 Meta engineers got reassigned to AI teams they didn't choose. Zuckerberg just admitted the bet 'hasn't progressed as quickly as expected.' The hidden risk of a big tech job right now is real. At a startup, at least the failure mode is yours.

[George Cheng @MrGeorgeCheng]

114 of 483 AI API models changed prices in March alone. OpenAI called their own pricing "accidental." Claude Sonnet 5 jumps 50% after August. If your margin depends on token costs staying fixed, open-source isn't ideology. It's unit economics.

[George Cheng @MrGeorgeCheng]

SaaS killed enterprise software licenses in the 2000s. AI agents are doing the same to SaaS now. $285B wiped from market caps. Seat pricing fell from 21% to 15% in 12 months. Price for outcomes, not access. The playbook just got rewritten.

[George Cheng @MrGeorgeCheng]

@DavidSacks Anthropic can’t be trusted.

[George Cheng @MrGeorgeCheng]

Gartner found ~130 vendors with real agentic AI. Thousands are selling it. The rest is RPA with a rebrand. Real test: give it something it wasn't trained for. Real agents adapt. Washed agents break.

[George Cheng @MrGeorgeCheng]

Tokenmaxxing was always a vanity metric. Uber burned its entire 2026 token budget in 4 months. Meta pulled its leaderboard. Microsoft cancelled subscriptions. Companies are finally asking what AI is actually producing, not just consuming.