Marc Andreessen accidentally told the truth about AI
In a Nutshell
Marc Andreessen claims AI makes engineers 20-50x more productive, but the reality is engineers now work 20-hour days chasing incremental prompt gains that rarely deliver shippable results. Companies are buying into the replacement narrative from OpenAI and Anthropic because it promises massive cost savings, yet AI is producing mostly technical debt that requires human sifting to extract value. The core issue is that managers are fixated on token volume instead of defining clear business objectives, and true progress still depends on human effort rather than agents.
These notes were generated by AI and may contain inaccuracies.
Marc Andreessen appeared on the Joe Rogan podcast where he claimed that AI is making engineers 20, 30, or 50 times more productive. He stated that people he knows are producing so much code that they cannot sleep because the opportunity cost is too high. Andreessen described engineers working 20-hour days voluntarily because their agents are constantly churning while they sleep.
The reality is that people are working 20-hour days because they are less productive and less efficient than before. The promise of one more prompt solving a long-standing problem creates a slot machine feeling. Engineers feel they are 88% complete and that one additional prompt will push them past the 98% mark, but each prompt only advances progress by 0.1%. The only way to win is to keep prompting 20 hours a day until something shippable emerges, which rarely happens.
The dystopia stems from the managerial and executive class pressuring employees to use AI tools. They push the technology saying it will make workers more productive without defining what productivity means or what business objectives to pursue. Managers and executives are making present decisions based on the future promised potential of the technology, expecting models to eventually become so good that organizations must be ready to take advantage.
This message is propagated by token salesmen at the top including Sam Altman and Dario Amodei. Anthropic and OpenAI are selling to enterprises, and the message enterprises want to hear is more productivity, more automation, and less need for fickle human beings. The narrative focuses on AI job displacement because that is the pitch that resonates with enterprise buyers. It is a trillion dollar sales pitch, and changing it to emphasize augmentation rather than replacement would signal to companies that they must pay for both humans and AI.
Companies stick with the enterprise human replacement pitch because it is the most profitable pitch in the history of capitalism. The next industrial revolution narrative convinces managers that AI will put their organization at the forefront of innovation. AI is presented as intelligence when it is actually a compelling parody of intelligence. Its usefulness has not been demonstrated from the output.
A manager's job is to tell employees what objectives to hit. The objective is not more tokens or having employees sit at the bottom of a token chute feeding tokens into their mouths. The objective is a business objective that managers must figure out. What tools employees use to accomplish the job hardly matters.
The LLM species has been discovered. It is like encountering an alien species on a foreign planet. These models are what they are. Giving them more tools does not make the base LLM suddenly smarter or round it out toward general intelligence. When new releases like Claude come out, this represents tool use by the same alien intelligence rather than fundamental improvement. AI is not getting smarter. It is the same base LLM technology learning to use different tools.
Reaching the next breakthrough toward AGI requires making a couple more scientific discoveries. Scientific discoveries of this magnitude happen on the order of once a century. Two more events on the scale of the fire and the wheel would be needed. Companies are running their revenue models assuming they will discover new continents, with the belief that the boats are getting fast enough to make this imminent.
Many companies are not figuring out how to make more money because making more money is hard. Layoffs represent an acknowledgement of that difficulty. Jason Fried of Basecamp compares bragging about token production to putting a finger on a camera shutter button and bragging about how many pictures are taken. Instead of capturing one, two, or three good photos, people take tens of thousands and must then review them all to find the ones that meet business objectives.
Tokens will soon be traded as commodities next to oil, with countries reporting GDP alongside token production and utilization metrics. Sam Altman is offering Y Combinator portfolio companies $2 million worth of tokens in exchange for real equity. Every token represents pressing down on the shutter button. Every token is technical debt. By the end, millions of tokens are produced that must be sifted through, mined, and reviewed. The job today is no longer software engineering. It is token refining. Every person is now a human sifter mining for gold in sparse tokens, working at the token refinery.
Before AI, engineers went from A to B with a business objective, fulfilled it by building the product, and typically completed work in less than eight hours. They left the codebase cleaner than they found it and felt productive. Now people work 20 hours and it is not enough because AI is not getting them to the objective.
The discussion concerns professional software engineering organizations building products used by millions or billions of users, such as iOS, Linux kernels, and banking apps. It does not apply to individuals shipping small dashboards, tools, or proof-of-concept projects. The token mania needs to calm down. Focus should return to business objectives.
Companies laying off 2,000 people to save a couple hundred million dollars achieve short-term number inflation, but multi-billion dollar companies could offset that with a new product. They are not doing so because figuring out what to build is hard. Marc Andreessen's claim of 50 times more demand for software is incorrect. There is not unmet demand for software. There is software that has no customers, no users, and no one willing to invest time figuring out how to use it. Breaking out of that pattern requires exceptional effort driven by people, not agents.
Keep Mo Bitar in your library
Save the videos and channels worth coming back to, and find them again in one place.
