After Code Got Cheap, the Most Expensive Person on the Team Started Saying LGTM
Gergely Orosz asks Satya Nadella: as generating code gets cheaper and verifying it gets harder, what happens to software engineering? The podcast discusses software factories and owning intelligence, and the comments section nails the new bottleneck.
Gergely Orosz's question in this podcast is very specific: as generating code gets cheaper and verifying code gets harder, what will software engineering become? The guest is Satya Nadella. The video is below, and the original tweet link is also kept.
Three parts worth watching.
1. Software factory
Satya gave the example of the Azure networking team: it manages more than 500 fiber network operators and recently built an agentic system to manage emails and processes, improving team efficiency. He called this kind of work the reality of AI transformation: brownfield, doing daily delivery while building a software factory to improve how the team operates.
Buying models and switching tools only changes the surface. What really needs to be dismantled are existing processes: who receives emails, who dispatches them, who checks them, and which steps can be handed to agents.
2. Owning intelligence
Satya warned companies not to get locked into a single model vendor. Harness, memory, and model should not be coupled; switching models should be easy. Open models can protect core IP, and private evals are used to decide which models to use and when to upgrade. He also said models are intelligent but unreliable, so it is best to treat them as an internal risk and build a separate system to verify agent behavior.
This is both architecture advice and procurement advice. Review capacity needs to be calculated in advance; don't wait until the generation side speeds up to discover that the verification side is queuing.
3. Microsoft's voluntary separation
Gergely asked directly why Microsoft is offering voluntary separation to most software engineers for the first time, and whether this might signal future layoffs. Satya's answer was diplomatic. In the podcast, Gergely restated what he heard in a more direct way.
The comments section is more direct than the interview. @Electra wrote, 'code got cheap, now the priciest person on the team says LGTM.' @Damon.Lab said, generating code has gotten cheaper, but the people who can spot what's wrong within 30 seconds haven't gotten cheaper—that's the new senior skill. @me hi said the bottleneck has moved from the keyboard to judgment. @jaan said ten agents write, but still one person reviews. @pirol offered new titles: senior diff reader, reporting to chief eval officer.
These replies all point to the same thing: the cost of generating code has fallen, but the cost of judgment and verification has not. If organizations only add people and agents to the generation side, the review side will become the new bottleneck.
Timestamps:
- 00:00 Intro
- 02:12 Microsoft's open-source shift
- 04:15 Building VS Code in public
- 05:55 Copilot, Cursor, and Claude Code
- 09:52 Decision engine
- 12:32 Verifying AI-generated code
- 15:39 Software factory
- 20:45 Bringing AI into existing products
- 24:06 Why adopting AI is different from adopting React
- 26:29 Changes in the software engineer role
- 29:25 Microsoft's voluntary separation plan
- 34:05 AI anxiety and role changes
- 36:35 Junior hiring
- 42:36 Open models
- 45:39 Owning your own AI infrastructure
- 51:21 Building trust in AI agents
- 53:48 Local models and hybrid intelligence
- 56:39 Microsoft's model strategy
- 59:26 Advice for engineers
- 1:01:21 Closing takeaways
Original tweet: https://x.com/GergelyOrosz/status/2122324513211728161
发布时间: 2026-11-16 22:27