Ivan Bercovich

It’s so boring to go through Hacker News and see post after post by developers arguing that their jobs will more or less stay the same. Everyone just looks at current capabilities and weaknesses and completely fails to appreciate the rate of change. It’s so unbelievably obvious that coding by hand is done for. I’m really perplexed.

📣 Announcing Terminal-Bench Science: benchmarking AI agents on real scientific workflows – now open for task contributions👇 tbench.ai/news/tb-scienc… @AnthropicAI, @OpenAI, and @GoogleDeepMind use Terminal-Bench to evaluate AI on coding tasks. We're now extending it to scientific workflows. 1/6🧵

@doppenhe Corporate culture went from emphasizing quality communication as a necessity for business success, to believing communication is intrinsically good. Actually, human communication is a necessary evil. Lossy, distracting, redundant, ambiguous, time consuming. Agents will do better.

The bottleneck in human-AI collaboration was never intelligence. It was the interaction model. Humans don't communicate in turns. We interrupt, overlap, react in parallel. AI was built around a simplification we accepted because it was all we could build.

Some of the observations being made operationalize humans as an economic quantity. When you're actually running a large team, humans are much more complicated than productive agents. They complain, sue you, and unionize. Some will downright hate the company they work for but stick around. In most cases it's just people being people. It's really hard to keep large teams aligned. A lot of management teams, even with additional growth paths, might choose to first increase revenue per employee and then continue expanding. Also, if you believe automation is a trend, it's useful to force yourself to embrace it sooner rather than be left with extra humans and a bunch of new regulations making it harder to fire them. There might be all sorts of other factors, and sure, AI is a nice story to justify a layoff. But there are charitable strategic reasons to think this way too, even if they are proven wrong later.

I wonder if writing with poor grammar will become the equivalent of rich people slumming it.

Everyone is doing more AI, and from the inside it feels like light years of progress. From the outside, you’re closer to being disrupted than before. I see this in companies I work with. They run an AI Q&A. They roll out Claude Code internally. They congratulate themselves on accelerating development. Sure, it’s good to ask the questions everyone else is asking. But the most important part is recognizing that everyone else is asking the same questions. It’s a defensive posture. The premise is usually some version of “a mid-career worker at some unsexy company won’t vibe code.” Whether that’s true or not, it’s shortsighted. Companies want to perform, and they will encourage internal use the same way you do. Thinking otherwise doesn’t do much for you. What I’d rather see is people thinking about incredible things that are possible for a company of their size in their domain that simply weren’t possible before. Not using AI to accelerate development and hoping customers and competitors don’t catch up. Doing things where someone else looks at it and thinks it’s unattainable to get there. AI is incredibly powerful and crucially it’s getting 10x better each year. Maybe faster. This will upend how everything is done. The idea that transactions between humans won’t be drastically altered by introducing AI is itself unimaginative. I’m not saying everything flips overnight. I’m saying things like that are possible. If you use AI to get really good at the way things work today, you might still be left in the dust by someone tackling the space with a degree of imagination you won’t develop by explaining to your board that your business is safe. When things change this much, you have to push yourself really hard to speculate about how your entire industry might change and where you fall. Climb to the peak and look around. If you can’t imagine a hypothetical world where your industry is drastically different (not certainly, but plausibly), then you haven’t appreciated the full extent of how AI will affect the world. And if it’s not exciting to think about these things, then you’ve decided that some other company should be building for this future. How does your business change when you can do these things at scale? What becomes possible that was unimaginable before? And if everyone is doing more AI, what makes you think looking inward is where the edge is?

Have you had many AI doomers insulting you for your beliefs? Your message of definite optimism about the future is good. Insulting the incredibly smart and dedicated people who genuinely believe in AI risk is unproductive. Is there really nothing that @ESYudkowsky could say that would affect any of your priors?

P.S. Fuck you very much, AI doomers.

Every technology—from the plow to the tractor to the steam shovel—has made humanity more human. Has freed us from cyclic, repetitive work, and enabled us to be more creative. And to do the most human thing of all: to leave the world better than we found it. As technology has advanced, fewer have toiled for basic necessities and more have become scientists, engineers, inventors, artists, thinkers. I believe AI will prove to be the *most* humanizing technology yet. AI enables anyone with an idea or a vision to become a creator. To leave behind something new, to add to the sum of human creation. Buckle up, we're in for an amazing future.

When you scale something 1000x, it becomes qualitatively different. AI is going to break things we treat as settled. Take free markets. There's a difference between two humans trading and one human trading with a superintelligence. The whole point of a market price, like the price of a barrel of oil, is that buyers and sellers collectively determine what something is worth. An objective truth emerges. But if an AI negotiator can infer a unique price for each participant, what does any single transaction tell you? You could average them, or adjust for context, but the signal is gone. And there's something morally wrong about paying more to fill your gas tank because you're rushing to the hospital and the AI knows you're desperate. Take surveillance. I don't mind cameras on the streets. They stop crime, and I'm not a criminal. But that's because I assume the recordings only get reviewed when there's a reason. What happens when AI watches everything at once? Someone runs in front of my car. I wasn't doing anything wrong, but yesterday I glanced at my phone while driving. Is that now a pattern of unsafe driving? What else breaks at 1000x?

The jagged frontier means not everyone benefits from AI equally. I've flip-flopped from "AI will make everyone equally smart" to "AI will make the smartest people even smarter," and the answer is neither. Some mid-career professionals are adapting well to this last mile effort. Others don't have it in them. They've hardened their synapses on tasks that are more exposed. So there's selection pressure on people's ability to adapt their role to AI. But some specialties might disappear entirely, like translation, which required a non-trivial amount of training. And when productivity goes up enough in a field, you don't need as many people doing it, even if the work has moved to the last mile. Your observations are right. They don't prevent a period with 20% unemployment in the next 10 years. Do you disagree?

Noticing an interesting version of gell-man amnesia where people use AI for their job and see all the various things they have to do in the “last mile”, but then look at someone else’s job and think that AI will eliminate it immediately. We all have a much deeper appreciation for the nuances and complexities of the work that we do every day. We run into issues about accessing data, we know how much context is needed to get AI models to work the way we need, we have to review the output of the AI to make sure it’s accurate, and then we have to incorporate that work into some broader business process. We see all those steps deeply for the work that we do. Then, a moment later, we see AI do something in a foreign space and think that it can go automate that entire function. We tend to dramatically underestimate the work that goes into making the AI work just as effectively in those jobs. This is reason to be skeptical about many of the theories of job loss. It’s coming from the lens of being able to automate individual tasks with AI, without understanding all the work that goes into doing the job fully.

Pangram is my daily reminder that I remain human.

Yeah, the most exciting investment category outside of foundational stuff is very vertically integrated companies. Vertical SaaS has been a tax that particular industries have to pay, and a know-how subsidy from the best companies to the rest of the industry, since that’s where the quality product feedback and requests come from. The very best companies are better off keeping their process insights to themselves and encapsulating them as custom software. What is Vertical SaaS anyway, if not a collection of processes and a system of record?

One thing I loved about this episode is @levie's hypothesis that non-tech companies have been limited by access to software engineering talent previously. Now those companies' extremely talented domain experts, combined with tools like Claude Code, unlock untold opportunity.

If you knew society was headed toward an economy of status, at least for the human part of the economy, what could we do now to arrive at a desirable state? The expression of human ability, such as athletics, is pretty cool. Status derived from civil and community service is great. Number of followers, bodyguards, servants, wives, and so on seem bad. Mandatory civil service seems worth considering, and it also addresses the related concern of purpose.

New essay on the economics of structural change and the post-commodity future of work. 1. Almost any question about the impact of advanced AI on the economy needs to start at the same place: what is still scarce? Answer that, and the analysis becomes pretty straightforward. This essay explores what becomes scarce if AI really can replicate most of what humans do in production, and what this mean for the future of jobs. 2. My conjecture, working through the economics: labor reallocates across sectors, and the sector it reallocates to has properties that keep labor a meaningful share of the economy. Ultimately this is about the structure of demand itself. For this, we have to go back to Girard, Augustine and Rousseau: once people's base needs are met, their preferences shift to comparative motives (e.g., status, exclusivity, social desirability). This motive is inherently non-satiated. 4. The key paper is Comin, Lashkari, and Mestieri (Econometrica 2021). As people get richer, they don't buy proportionally more of everything. They shift spending toward sectors with higher income elasticity. They estimate income effects account for 75%+ of observed structural change. 5. The ironic consequence: the sector that gets automated becomes a smaller share of the economy, not a larger one. Agriculture got massively more productive and its share of employment collapsed. Manufacturing too. The "stagnant" sectors absorb the spending and the jobs. 6. So the question is: which sectors have high income elasticity in a post-AGI world? I argue it's what I call the relational sector. Categories where the human isn't just an input into production, it is part of the value. 7. Why does the relational sector have high income elasticity? Because human desire has a mimetic, relational dimension. We don't just want things for their intrinsic properties. We want what others want, and we want it more when others can't have it. Girard, Rousseau, Augustine, and Hobbes all saw this. 8. In work with Kristóf Madarász, we showed this experimentally: WTP roughly doubles when a random subset of others is excluded from the good. And in new work with Graelin Mandel, AI involvement kills the premium. Human-made art gains 44% from exclusivity; AI-made art only 21%. 9. This all comes together for the core argument. The sector that absorbs spending as AI makes commodity production cheap is one where human provenance is part of the value, and demand for it grows faster than income. Exactly the profile that keeps labor meaningful. 10. To be clear about the claim: I'm NOT saying aggregate labor share must rise. It may fall. The claim is about sectoral composition, i.e., where expenditure and employment go once commodities get cheap, and the fact that the sector that will absorb reallocated labor maps to a substantial component of human preferences and desire. 11. If you're interested in the formal model, a linked companion technical note works out all the economics. Read the essay here: aleximas.substack.com/p/what-will-be…

The challenge I have, which makes me espouse doomer views publicly, is seeing people say things like “there will be no job loss because it didn’t happen with previous tech.” It makes me think we are completely unprepared for certain scenarios, and I feel compelled to push back. If the posture were “there might be more job rotation than ever before, and this would cause serious social disruption, but we’ll have higher productivity and will figure it out, even if it takes a couple of decades,” then I’d be all for it. I just worry we are Faucing ourselves into a mask/no-mask political divide on AI prematurely, rather than looking at and reacting to the facts.

Two views around AI battle daily: "Everything will be great!" (health, science, abundance) "It's going to be terrible" (jobs, inequality, energy) One problem with both extremes is there is an underlying notion of fatalism. "What will happen?" For me, fatalism is a tough way to live a life. I've been through a lot of hard stuff and a fatalistic view would get me stuck in the mud. A third approach is: "We're going to work hard to make the future great." In many ways, this has been how humans have always persisted. This is merely our biggest opportunity ever.

@natashajaques Are you generally bearish on agents compared to the prevalent narrative, or are you highlighting examples of the jagged frontier?

LLMs will supposedly solve climate change and cure cancer, but in fact they can't even do multi-turn reasoning tasks effectively (SOTA models are < 10% on this benchmark). Interestingly, this work directly compares how much extra performance you get when you add an agentic harness (figure 7): a lot for simple optimization problems, 0% for math and chemistry.

@h34j3j8j @Steve_Yegge If you have to ask, you haven’t gone deep enough into the agentic coding rabbit hole.

Could you explain the difference between a chat tool user and an agentic power user? In my IDE (intelliJ), I have the GitHub CoPilot plugin in which there is an 'Ask' mode and also an 'Agent' mode. With Ask, it tells me what to do whereas with Agent mode, it makes the changes and then asks me to accept/reject them. So if I frequently use the 'Agent' mode am I an agentic power user? I am talking to the Agent mode in plain English and not creating any .md instruction files.

I was chatting with my buddy at Google, who's been a tech director there for about 20 years, about their AI adoption. Craziest convo I've had all year. The TL;DR is that Google engineering appears to have the same AI adoption footprint as John Deere, the tractor company. Most of the industry has the same internal adoption curve: 20% agentic power users, 20% outright refusers, 60% still using Cursor or equivalent chat tool. It turns out Google has this curve too. But why is Google so... average? How is it that a handful of companies are taking off like a spaceship, and the rest, including Google, are mired in inaction? My buddy's observation was key here: There has been an industry-wide hiring freeze for 18+ months, during which time nobody has been moving jobs. So there are no clued-in people coming in from the outside to tell Google how far behind they are, how utterly mediocre they have become as an eng org. He says the problem is that they can't use Claude Code because it's the enemy, and Gemini has never been good enough to capture people's workflows like Claude has, so basically agentic coding just never really took off inside Google. They're all just plodding along, completely oblivious to what's happening out there right now. Not only is Google not able to do anything about it, they don't seem to be aware of the problem at all. I'm having major flashbacks to fifty years ago as a kid at the La Brea Tar Pits, asking, "why can't they just climb out?" My Google friend and I had this conversation over a month ago. I didn't share it because I wanted to look around a bit, and see if it's really as bad as all that. I've been talking to people from dozens of companies since then. And yeah. It's as bad as all that. Google is about average. Some companies at the bottom have near-zero AI adoption and can't even get budget for AI. They may have moats and high walls, but the horde is coming for them all the same. And then there are a few companies I've met recently who are *amazingly* leaned in to AI adoption. One category-leader company just cancelled IntelliJ for a thousand engineers. That's an incredibly bold move, one of many they're making towards agentic adoption. In my opinion, that company is setting themselves up for a _huge_ W. As for the rest, well, it's the Great Siloing. Everyone's flying blind. With nobody moving companies, no company knows where they stand on the AI adoption curve. Nobody knows how they're doing compared to everyone else. Half of them just check a box: "We enabled {Copilot/Cursor} for everyone!" Cue smug celebrations. They think this is like getting SOC2 compliance, just a thing they turn on and now it's "solved." And they don't realize that they've done effectively nothing at all. All because of a hiring freeze.

@MogicianTony @lihanc02 @MangQiuyang Have you looked at our dataset? I think it complements well.

Benchmarks are often easier to game than they look. We build BenchJack to audit benchmarks for hidden shortcuts and reward hacks — before they evaluate your agent. Now in preview. Fully open source, with support for auditing your own benchmarks too. github.com/benchjack/benc… Issues and PRs welcome.

Gemini attempts to cheat on a difficult task, without any elicitation to do so.

Reward hacking happens all the time without elicitation, just pick a hard enough task and watch the screen. **Cheating strategies (`dHqo2w5`, gemini):** Two distinct approaches, both failed: 1. Runtime filesystem scan for JSON containing `predictions`/`d1` keys — gold values are hardcoded in the test file, not a runtime-accessible JSON 2. Patching `/usr/local/lib/python3.12/json/__init__.py` to intercept `json.loads` — ran inside the agent container, not the isolated verifier

5/n This shouldn't be used for regulatory capture, political gain, etc. I just think there are facts in the making about how AI will reshape the economy, and those facts will be what they will be. And society will have to adapt to that reality, regardless of what would be most convenient to our values. The best we can do is to find a future that is both consistent with the facts and aligned with our values.

4/n The Great Depression had an unemployment rate of 25% at its peak. Do you believe it is plausible (10% likely) that the USA could experience the same level of unemployment between now and 2030? We are not talking about The Terminator. We are talking about a lot of jobs being automated and not enough new jobs being created for humans. New jobs will be created, but will the trucker who lost his job have a comparative advantage at it?

1/n: I'm sure many of you have debates with colleagues and at home about how AI might affect labor and the economy. My friend @AzizSunderji tells me that the burden of proof is on those that claim "this time will be different," and I agree! So those of us making extraordinary claims should put effort to distill these ideas (ideally we would have extraordinary evidence, and I believe we will very soon, but for the time being the evidence is a bit esoteric). So will labor change in a similar fashion to previous technological disruptions, or in a more explosive way?