Are developers happy yet? Unpacking the 2025 Developer Survey | Stack Overflow’s Erin Yepis
Plus, a lobotomized LLM, charting the AI boom, & preventing burnout during the holidays
After hitting a low point last year, developer job satisfaction is officially on the rise. Erin Yepis returns to the show to unpack the 2025 Stack Overflow Developer Survey, analyzing how autonomy and compensation are driving this recovery. We also cover the happiness gap between senior and junior engineers, the surprising drop in trust for AI tools, and why vibe coding is failing to catch on with professional engineers.
1. DevAI: The vanity metric trap
AI dramatically accelerates code generation, but raw code output is a vanity metric; if you generate code faster than you can validate it, the bottleneck simply shifts to testing and deployment. As noted in this piece, AI-assisted codebases decay from greenfield to legacy faster than ever, creating unmaintainable systems in record time. Sustainable adoption requires experienced engineers who know when to override the bot, ensuring you don’t just build technical debt at 100x speed.
Read: DevAI: Beyond Hype and Denial
2. Quantum physicists just lobotomized an LLM
A Spanish firm specializing in quantum-inspired AI techniques has successfully compressed the DeepSeek R1 model by over 50%, shrinking its tensor size while revealing the censorship pathways built into the system. While the censorship angle is fascinating, the engineering implication is bigger: it shows how easily LLMs can be manipulated without training new models from scratch. We are entering an era where AI systems are becoming transparent targets for extraction. The same tools used to build intelligence layers can be used to reverse engineer, repurpose, and rebuild them with surprising ease.
Read: Quantum physicists have shrunk and “de-censored” DeepSeek R1
3. Adoption is easy. Trust is hard.
It is easy to buy Copilot seats. It is much harder to know if your team actually trusts the suggestions enough to move faster. High adoption with low acceptance rates is a red flag. It signals that your engineers are fighting the tool rather than using it. LinearB’s new dashboards expose these trust patterns by unifying data from Copilot and Cursor and mapping it directly to delivery impact. Stop treating AI as a sunk cost and start treating it as a measurable engineering asset.
Read: Measuring the impact of Copilot and Cursor on engineering productivity
4. 16 charts that define the boom
Capital expenditures on AI have gone vertical in the last 12 months, dwarfing the peak annual spending of historical moonshots like the Manhattan Project or the Apollo program. While power consumption is a major headline, the macro data suggests it is actually outpaced by the aggregate energy demands of electric vehicles. However, the most startling chart may belong to OpenAI’s cash flow projections: they are currently forecasting a peak negative cash flow of $40 billion in losses in 2028. Have we reached peak hype yet?
Read: 16 Charts that explain the AI boom
5. The 6-month “shaky” period
Friend of the show Kelly Vaughn argues that the “shakiness” you feel in a new role -that imposter syndrome hitting months in - isn’t failure, but a standard psychological experience that can persist for up to six months. We need to normalize this transition period rather than fighting it. With the holidays approaching, it serves as a timely reminder for leaders: ensure your teams are actually resting and recharging, rather than just burning out while trying to prove themselves in a new environment.
There is a hollow irony in celebrating "recovery" while we simultaneously automate the decay of our codebases. We are generating legacy systems at 100x speed, trading the quiet understanding of the builder for the frantic validation of the reviewer.
Do you feel like you are still crafting the solution, or just cleaning up after the machine?
Re the lead-in line "While power consumption is a major headline, the macro data suggests it is actually outpaced by the aggregate energy demands of electric vehicles" isn't quite what the underlying IEA report says. The chart showing absolute TWh growth understates data centers' significance—the IEA report it draws from notes data center demand grows at a faster rate (4x faster) than overall electricity growth. The IEA presents multiple scenarios with higher demand by 2035. More importantly, data centers cluster geographically—nearly half of U.S. capacity is in five regional clusters—creating local grid strain (eg, seasonal that national averages obscure.