When the overall economy grows, the demand for human labor often grows with it — it doesn’t shrink.
That has been the story of 200+ years of industrialization: new technology eliminates some jobs, but it also creates new industries, new demand, and new kinds of work.
We heard the same panic about radiologists. In 2016, Geoffrey Hinton famously suggested we should stop training radiologists because AI would outperform them. Yet in 2026, we need more radiologists, not fewer. The job is changing, not disappearing.
You even see a similar dynamic with immigration. Immigrants don’t just “take jobs”; they also create demand, start businesses, pay taxes, and expand the market. Remove them, and the economy often shrinks — meaning fewer jobs overall, not more.
TL;DR: AI is not simply “coming for your job.” Yes, the nature of work will change. We no longer employ “human calculators,” but society didn’t run out of work. We created better, more productive jobs than doing arithmetic by hand all day.
1. Chat, being 3 yr old, is a fairly mature and solved problem today. Top companies aren't even talking about it anymore! Gemma 31B does it amazingly well (for $0.4/1M token output). Practically every near-SoTA and SoTA model does simple "chat-like" QA amazingly well -- summarization, basic question answering, single- or few-step search.
2. Tasks -- or knowledge work on a computer -- are the new frontier. Computers have become competent only recently, and only for some of the tasks so far. I'd guess another 2-3 yr development cycle, after which "el cheapo" models will be virtually indistinguishable from SoTA.
As tasks are the new game in town, AI labs can still charge a premium for it. That premium has disappeared already for chat; most users cannot tell 99% correct answer from 95% correct answer; nor do they always wish for maximum accuracy.
3. What comes after Tasks? I think today's AI startups should figure that one out and solve it before everyone else.
For those who missed: Gemini coding and agentic capabilities have been lagging the sota models (Opus mostly) since Dec 2026. If you're a co-lead and your model is underperforming there has to be some consequences. I don't know as a fact if this has anything to do with Noam's departure, but work performance is never about past successes.
I owned a model 3 and two model Ss. I largely regard model S as the best sedan ever made (in less than 100k USD luxury price range). No, model 3 doesn't even come close. Plus you can buy a 3yr old model S "like new" at the price of a new model 3.
Tesla is committing suicide here by eliminating the best sedan ever and by committing to an idea (taxis) that mainly serve the low end market.
Message to Musk: people like to own things. Only the low end segment don't mind sharing their means of transportation. High end won't share. Biggest profits in tech are in the high end market as Apple and Samsung have repeatedly shown.
How to read this brilliant blog post by Gary Marcus?
- Replace all the occurances of "LLM" by "human";
- Replace all the occurences of "Scaling" by "additional education".
Voila! You get an article that actually makes sense, plus you'll get a better sense of where the technology is -- these models are behaving much like humans across many tasks. They aren't perfect. But they are getting better everyday, and are quite useful.
Thank you AI developers and researchers for making progress everyday! No "thank you" to people like Gary Marcus who'd be called a "perma bear" in financial parlance.
Despite consensus forecasts predicting a slowdown to 1.8% growth, the US economy appears poised for acceleration in 2026 driven by a potent combination of aggressive fiscal and monetary loosening. Treasury Secretary Scott Bessent’s optimism is underpinned by the implementation of the "One Big Beautiful Bill Act," which delivers retroactive tax cuts, alongside a rebound in government spending following a record 43-day shutdown and the potential for the Supreme Court to invalidate certain tariffs, resulting in significant corporate refunds. This fiscal stimulus coincides with the Federal Reserve’s pivot to lower interest rates—a trend likely to intensify as President Trump seeks to appoint dovish leadership to the central bank. While this synchronized stimulus supports bullish stock market projections and complements favorable global conditions like low oil prices, it carries significant risks of reigniting inflation and spiking long-term bond yields; nevertheless, the absence of immediate shocks suggests the economy has ample scope to outperform expectations.
By late 2025, Boston’s Kendall Square biotech hub faces a severe downturn marked by a "biotech winter" of plummeting venture capital and soaring lab vacancies. Caused by high interest rates, domestic policy uncertainties, and intensifying global competition, this crisis has triggered a significant talent exodus, leaving recent PhD graduates overqualified and underemployed as companies freeze hiring to cut costs. The contraction has stalled critical medical research and threatened Boston’s economic stability, though industry leaders remain cautiously optimistic that adaptation strategies—such as AI integration and renewed merger activity—could spark a recovery by 2026.
This reads like a mid-life crisis. A few rebuttals:
1. Yes, humans cause enormous harm. That’s not new, and it’s not something a single technology wave created. No amount of recycling or moral posturing changes the underlying reality that life on Earth operates under competitive, extractive pressures. Instead of fighting it, maybe try to accept it and make progress in other ways?
2. LLMs will almost certainly deliver broad, tangible benefits to ordinary people over time; just as previous waves of computing did. The Industrial Revolution was dirty, unfair, and often brutal, yet it still lifted billions out of extreme poverty in the long run. Modern computing followed the same pattern. LLMs are a mere continuation of this trend.
Concerns about attribution, compensation, and energy use are reasonable to discuss, but framing them as proof that the entire trajectory is immoral or doomed misses the larger picture. If history is any guide, the net human benefit will vastly outweigh the costs, even if the transition is messy and imperfect.
The distinction Karpathy draws between "growing animals" and "summoning ghosts" via RLVR is the mental model I didn't know I needed to explain the current state of jagged intelligence. It perfectly articulates why trust in benchmarks is collapsing; we aren't creating generally adaptive survivors, but rather over-optimizing specific pockets of the embedding space against verifiable rewards.
I’m also sold on his take on "vibe coding" leading to ephemeral software; the idea of spinning up a custom, one-off tokenizer or app just to debug a single issue, and then deleting it, feels like a real shift.
The author is conflating a financial correction with a technological failure.
I agree that the economics of GenAI are currently upside down. The CapEx spend is eye-watering, and the path to profitability for the foundational model providers is still hazy. We are almost certainly in an age of inflated-expectations hype-cycle peak that will self-correct, and yes, "winter is harsh on tulips".
However, the claim that the technology itself is a failure is objectively disconnected from reality. Unlike crypto or VR (in their hype cycles), LLMs found immediate, massive product-market fit. I use K-means clustering and logistic regression every day; they aren't AGI either, but they aren't failures.
If 95% of corporate AI projects fail, it's not because the tech is broken; it's because middle management is aspiring to replace humans with a terminal-bound chatbot instead of giving workers an AI companion. The tech isn't going away, even if AI valuations might be questioned in the short term.
AI-text detection software is BS. Let me explain why.
Many of us use AI to not write text, but re-write text. My favorite prompt: "Write this better." In other words, AI is often used to fix awkward phrasing, poor flow, bad english, bad grammar etc.
It's very unlikely that an author or reviewer purely relies on AI written text, with none of their original ideas incorporated.
As AI detectors cannot tell rewrites from AI-incepted writing, it's fair to call them BS.
What kind of world do you live in? Actually Google ads tend to be some of the highest ROI for the advertiser and most likely to be beneficial for the user. Vs the pure junk ads that aren't personalized, and just banner ads that have zero relationship to me. Google Ads is the enabler of free internet. I for one am thankful to them. Else you end up paying for NYT, Washinton Post, Information etc -- virtually for any high quality web site (including Search).
Lowering LDL cholesterol is arguably the most evidence-backed longevity intervention available today. Mendelian randomization studies suggest that each standard deviation of lifelong LDL reduction translates to roughly +1.2 years of additional lifespan, implying ~+2.4 to +3.6 years from sustained, meaningful lowering alone.
Pair this with tight blood-pressure control (aim systolic <130 mmHg) and a healthy BMI—every incremental improvement helps. Together, LDL, BP, and BMI form the most potent triad of interventions most people can implement now and expect to see substantial benefits 20–40 years down the line.
I recently downloaded about 10 years of monthly price returns for QQQ, TQQQ, NVDA, GBTC, and a few others. Then I asked ChatGPT and Gemini (separately) to find the portfolio that maximizes an adjusted CAGR — roughly, mean return minus ½ × standard deviation².
Result: 70% NVDA, 30% GBTC (Bitcoin), and 0% QQQ or TQQQ.
Honestly, not a bad mix — especially for a small, high-risk slice of your portfolio.
Next, I compared TQQQ (Triple Qs) vs. QQQ using the same 10-year monthly data. The optimizer picked 100% TQQQ, which again makes sense if you’re doing this in a tax-advantaged account like a 401(k) or IRA and only with money you’re willing to take some risk on.
Then I expanded the dataset — 55 years of returns across major asset classes (S&P 500, gold, short- and long-term Treasuries, corporate bonds, real estate, etc.) — and asked for the optimal portfolio.
The winner: ~85% S&P 500, 15% gold, though 75/25 gives nearly the same return with a better Sharpe ratio.
A few quick takeaways:
Gold → GLDM ETF is the best vehicle.
QQQ → QQQM or TQQQ are the best versions.
And if you’re feeling adventurous: 70% NVDA, 30% IBIT (Bitcoin) isn’t crazy.
For what it’s worth, I’ve been running 75% stocks / 25% gold for a while now, but I’m thinking of carving out ~10% of the stock portion for a more aggressive tilt: TQQQ (6%), NVDA (2%), IBIT (1%) — because why not?
1. I find Gemini 2.5 Pro's text very easy and smooth to read. Whereas GPT5 thinking is often too terse, and has a weird writing style.
2. GPT5 thinking tends to do better with i) trick questions ii) puzzles iii) queries that involve search plus citations.
3. Gemini deep research is pretty good -- somewhat long reports, but almost always quite informative with unique insights.
4. Gemini 2.5 pro is favored in side by side comparisons (LMsys) whereas trick question benchmarks slightly favor GPT5 Thinking (livebench.ai).
5. Overall, I use both, usually simulatenously in two separate tabs. Then pick and choose the better response.
If I were forced to choose one model only, that'd be GPT5 today. But the choice was Gemini 2.5 Pro when it first came out. Next week it might go back to Gemini 3.0 Pro.
I’m a big advocate of branchless programming — keeping configurations to a minimum and maintaining as much linear flow as possible, with little to no cfg-driven branching.
Why? I once took over a massive statistics codebase with hundreds of configuration variables. That meant, in theory, upwards of 2^100 possible execution paths — a combinatorial explosion that turned testing into a nightmare. After I linearized the system, removing the exponential branching and reducing it to a straightforward flow, things became dramatically simpler. What had once taken years to stabilize, messy codebase, became easy to reason about and, in practice, guaranteed bug-free.
Some people dismissed the result as “monolithic,” which is a meaningless label if you think about it. Yes, the code did one thing and only one thing —- but it did that thing perfectly, every single time. It wasn’t pretending to be a bloated, half-tested “jack of all trades” statistics library with hidden modes and brittle edge cases.
I’m proud of writing branchless (or “monolithic” code if you prefer). To me, it’s a hallmark of programming maturity -- choosing correctness and clarity over endless configurability, complexity and hidden modes.
He's a very smart and imaginative guy. The most imaginative people sometimes aren't the best at predicting the future (top 1), but rather good at predicting possibilities (top N).
I've seen this happen before. President ordering people to be sued, and even jailed. There's a name for this kind of system and it isn't congressional democracy.
That has been the story of 200+ years of industrialization: new technology eliminates some jobs, but it also creates new industries, new demand, and new kinds of work.
We heard the same panic about radiologists. In 2016, Geoffrey Hinton famously suggested we should stop training radiologists because AI would outperform them. Yet in 2026, we need more radiologists, not fewer. The job is changing, not disappearing.
You even see a similar dynamic with immigration. Immigrants don’t just “take jobs”; they also create demand, start businesses, pay taxes, and expand the market. Remove them, and the economy often shrinks — meaning fewer jobs overall, not more.
TL;DR: AI is not simply “coming for your job.” Yes, the nature of work will change. We no longer employ “human calculators,” but society didn’t run out of work. We created better, more productive jobs than doing arithmetic by hand all day.