What this covers

This is a 21-minute onstage conversation from the April 2023 MIT event, in which Sam Altman fields questions from the moderators and the audience. It begins with advice to entrepreneurs building on GPT-4. He argues that platform shifts raise the premium on ordinary business discipline, and he looks back at the App Store era, when thin, exploitative apps rose and collapsed while companies that kept up product quality and user focus lasted more than a decade. His most contested piece of guidance is to let tactics become strategy: watch what works with users, do more of it, and don't commit early to a fixed plan.

The middle of the conversation turns to scaling. An unidentified questioner presses on parameter count and on how much prompt engineering and in-context learning can do without retraining. The exchange separates continued growth in model size from the surprising emergent behavior of LLMs). Altman says that a new foundation model company would now have to catch up to trillion-parameter systems, or hope for a paradigm that makes them cheaper. Audience anecdotes about productivity run from a venture fund seeing code-generation gains of 5% to 20x to a claimed 10–100x lift for authors.

The later portion covers OpenAI and safety. Altman describes more than six months of audits and red-teaming between finishing GPT-4 training and releasing it. He says OpenAI is working on other safety issues that the open letter left out, and that its public openness is worth the cost of sometimes being wrong. He rejects one-day or one-month takeoff scenarios, partly because data centers and power take years to build. He also says LLMs already help build the next generation of AI but that humans will drive the loop for a long time. The talk closes with his own habit of choosing long-horizon, capital-intensive technology problems because few others try them.

Sharpest takeaway

Sam Altman advises entrepreneurs building on GPT-4 and upcoming models that enduring business success requires traditional fundamentals—product-market fit, user obsession, and long-term strategy—rather than relying solely on the platform shift, and that capabilities rather than parameter counts should guide development priorities.

  • App Store analogy: lightweight exploitative companies rose and fell while traditional business discipline endured 15 years
  • Focus on capabilities over parameter counts parallels the abandoned gigahertz race in chips
  • Tactics-to-strategy approach: tight feedback loops on user observation outperform speculation

The claims · ranked41 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.80

During a platform shift like GPT-4's emergence, building a product people love, staying close to users, and maintaining long-term durable business strategy becomes more important, not less important, compared to times of technological stability.

normativehigh valueestablishednovelty 2/4durability 4/4· Sam Altman

“nothing lets you off the hook for building a product that people love for being like very close to your users fulfilling their needs for thinking about a long-term durable business strategy like that's actually probably only more important during a platform shift not not less”

0.74

Only three things matter for company success during platform shifts: (1) building a product people love, (2) staying very close to users and fulfilling their needs, (3) thinking about long-term durable business strategy—and the technology is simply an enabler.

normativehigh valueestablishednovelty 1/4durability 4/4· Sam Altman

“I think there's always more that stays the same about how to make a good company than what changes the the and a lot of people whenever there's like a new platform shift like this thing just because they're using the platform like that's what's going to guide business strategy it doesn't nothing lets you off the hook for building a product that people love for being like very close to your users fulfilling their needs for thinking about a long-term durable business strategy”

0.70

Long-term, durable business strategy requires building sustainable competitive advantages that compound over time, and using the technology as an enabler but not letting it become the strategy itself.

normativehigh valueestablishednovelty 1/4durability 4/4· Sam Altman

“you definitely want to be in that ladder category and the tech the technology is just like this is a new enabler um but what you have to do as a company is like build a great company that has a long-term compounding strategic advantage”

0.69

The App Store platform shift (circa 2008) produced lightweight companies with exploitative mechanics that had meteoric rises and falls, while companies that applied traditional business discipline—product quality, user focus, long-term strategy—endured successfully for 15+ years afterward.

factualhigh valueestablishednovelty 1/4durability 3/4· Sam Altman

“there were a ton of companies that built very lightweight things with I don't want to call them like exploitative mechanics but just like you know it was it was not something durable and those companies had incredible meteoric Rises and falls and then the companies that really like did all the normal things you're supposed to do to build a great business endured for the last 15 years”

0.68

For new platform shifts like GPT-4, the optimal strategy is to let tactics become strategy: tightly observe what works with users, do more of it, and less of what doesn't work, rather than committing to a fixed strategic plan from the outset.

normativehigh valuecontestednovelty 2/4durability 3/4· Sam Altman

“I think in new areas like this one of the right approaches is to let tactics become strategy instead of the other way around and you know I have my ideas I'm sure you all have your ideas maybe we'll be mostly right we'll be wrong in some ways and and even the details of how we're right will be wrong about um the I think you never want to lose sight of vision and focus on the long term but a very tight feedback loop of paying attention to what is working and what is not working and doing more of the stuff that's working and less of the stuff that's not working and just very very careful user observation can go super far”

0.64

As the underlying substrate of LLMs gets faster and more capable, these algorithms will do surprising things on relatively short time scales because they scale non-linearly—surprise scales with capability, not linearly with computational input.

causalhigh valuecontestednovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“so this class of algorithm then if the underlying substrate is getting faster and faster and faster it's going to do surprising things on a relatively short time scale”

0.62

Speculation about the next breakthrough use cases for AI is less valuable than putting something into the world, deeply observing what happens, and being responsive to user feedback.

normativehigh valuecontestednovelty 1/4durability 3/4· Sam Altman

“I can speculate on ideas you all can speculate in ideas none of that will be as valuable as putting something out there and really deeply understanding what's happening and being responsive to it”

0.61

There is a meaningful distinction between having scaling continue at large parameter counts (moving from 1 trillion to 10 trillion to 10 quadrillion parameters) versus achieving surprising emergent capabilities where the algorithm does things that weren't anticipated purely through scale.

definitionhigh valuecontestednovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“but one thing that's absolutely unique about this class of algorithm versus anything I've ever seen before uh is that it it surprises you with raw horsepower regardless of whether you measure it in parameter count or some other way uh it it does things that you didn't anticipate purely by putting more horsepower behind it”

0.61

The questioner notes that the ability to achieve with prompt engineering and in-context learning grows disproportionately with model size (e.g., going from 1B to 175B to 1T parameters yields super-linear improvements in what you can do without retraining).

factualhigh valuecontestednovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“regardless of the specifics the trend is toward as the models are getting bigger and bigger and bigger so you go from one trillion to 10 trillion parameters the amount you can achieve with just changing uh prompt engineering or changing the the tokens that are feeding into it is growing disproportionately to the model size does that sound right”

0.61

Language models' ability to scale to surprising new capabilities is analogous to but distinct from simple hardware speed increases, in that new qualities emerge rather than just quantitative improvements in the same function.

definitionhigh valuecontestednovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“if you have a spreadsheet you coated it up you run it on a computer that's 10 000 times faster it doesn't really surprise you it's it's nice and responsive it's still a spreadsheet whereas this class of algorithm does things that it just couldn't do before”

0.55

OpenAI's goal is to get the world to engage with AI, think about it, gradually update understanding, and build new institutions or adapt existing ones to figure out what future humanity wants in an AI-enabled world.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“so we're going to keep doing that because we think it's better than any alternative and a big part of our goal at open AI is to like get the world to engage with this and think about it and and gradually update and build new institutions or adapt our existing institutions to be able to figure out what the future we all want is”

0.55

Starting a foundation model company from scratch now would require catching up to existing trillion-parameter models, or relying on the discovery of a new paradigm that makes current approaches simpler or less expensive, which would be beneficial for the world but not for existing AI companies.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“so then so then somebody trying to start from scratch somebody trying to start from scratch you know is trying to catch up to something or maybe or maybe we're all being incredibly dumb and we're missing one big idea and all of this is not as hard or expensive as we think and there will be a totally new paradigm that obsoletes us which would be great and not great for us but like great for the world”

0.52

OpenAI's philosophy is to remain open and honest in public communications, even though this means sometimes saying things that turn out to be wrong, because the technology's impact on everyone is so significant that public engagement and discussion is worth the trade-off of occasional public embarrassment.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“we certainly plan to continue doing that um the trade-off is like we say dumb stuff sometimes you know stuff that turns out to be totally wrong and I think a lot of other companies don't want to say something until they're sure it's right um but I think this technology is going to so impact all of us that we believe that engaging everyone in the discussion putting these systems out into the world deeply imperfect though they are in their current state so that people get to experience them think about them understand the upsides and the downsides it's worth the trade-off even though we do tend to embarrass ourselves in public and have to change our minds with new data frequently”

0.52

Within OpenAI, LLMs are already being used as a force multiplier in the creation of the next iteration of AI systems, but humans will continue to drive the improvement loop for a long time rather than AI achieving full recursive self-improvement.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“I think that it is going to be a much fuzzier boundary for you know getting to self-improvement or or not um I think what will happen is that more and more of the Improvement Loop will be aided by AIS but humans will still be driving it um and it's going to go like that for a long time”

0.52

In entrepreneurship, predicting what the next real-world, society-benefiting use case will be that hits the tipping point on the LLM capability curve is a critical strategic question for the entrepreneurs in the room.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“so I think one of the things the people in the room need to predict is okay what is the next real world society benefiting use case that hits that Tipping Point on this curve”

0.52

OpenAI spent more than six months after completing GPT-4 training before releasing it to conduct safety studies, external audits, red-teaming, and mitigation efforts.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“we spent more than six months after we finished training gpt4 before we released it so taking the time to really study the safety of a model to get external audits external red teamers um to to really try to understand what's going on and mitigate as much as you can that's important”

0.52

OpenAI is doing other things on top of GPT-4 that have all sorts of important safety issues that were left out of the open letter's concerns.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“but we are doing other things on top of gpt4 that I think have all sorts of safety issues that are important to address and we're totally left out of the letter”

0.52

Altman does not believe in rapid (one day or one month) takeoff scenarios for AI self-improvement, partly because infrastructure (data centers, power, concrete) takes a very long time to build even if the technology is understood.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“there's like a whole bunch of other things that I have never believed in the like one day or one month takeoff um for a bunch of reasons but like one of which is how incredibly long it takes to build new data centers bigger data centers like even if we knew how to do it right now just like waiting for the concrete to dry getting the power into the building this stuff takes a while”

0.52

The rate of change in the world will increase continuously and accelerate as humans get better and better tools, but most sci-fi predictions about how that plays out (including singularity scenarios) end up being inaccurate.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“but I think what will happen is humans will be more and more augmented and be able to do things in the world faster and faster and it will not work out like it will not somehow like most of these things don't end up working out quite like the Sci-Fi books and neither will this one but the rate of change in the world will increase forever more from here as humans get better and better tools”

0.52

OpenAI was in development for seven years before ChatGPT, and the company was sweating every detail of technology and product throughout that period, which is why ChatGPT succeeded where others had not.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“I mean we've been working on it for so long that it's like you kind of know with gradually increasing confidence that it's it's really going to work but this is you know we've been doing the company for seven years um these things I would say bye and like in terms of why it worked when others haven't it's just because we've like been on the grind sweating every detail for a long time and most people aren't willing to do that”

0.52

GPT-4 can now generate book text so well that an author can hand the framework and concept to GPT-4, which fills in the words beautifully, creating a roughly 10–100x force multiplier for author productivity compared to what was possible with GPT-2.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“you have to give it the framework you have to you know you're still effectively writing the concept but it's filling in the words just beautifully and so as an author that could be a force multiplier of something like 10 100 and it just enables an author to be that much more powerful”

0.52

Altman's only consistent strategic approach is to pursue long time horizon, capital intensive, difficult technology endeavors, and he does not believe he is particularly good at them—he does them because few others try, resulting in low competition.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“the only thing I know how to do I think is this one strategy again and again which is very long time Horizon Capital intensive difficult technology bits and I don't even think I'm particularly good at those I just think not many people to try them so there's very little competition which is nice I mean I don't have a lot of competition”

0.52

The moderator suggests that Altman is at a historic inflection point where he will be a centerpiece of history books no matter how AGI develops, and that if AGI arrives in the near term, this will divide human history into a period before AGI and a completely different period after.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“if AGI is is in the near-term future uh then we're right now at this inflection point where human history has a period of time up till AGI and then obviously has a completely different history from here forward so it seems to me that at this stage you're going to be a centerpiece of the history books no matter how this evolves”

0.52

The questioner's venture fund is experiencing a wide range of productivity lifts from LLM-assisted code generation, ranging from 5% to 20x improvement depending on the type of code, context requirements, and system tuning.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“a lot of the companies that I'm an investor in or have been co-founder of are starting to use llms for cogeneration and it is interesting very wide range of lifts or Improvement in the performance of an engineer ranging from about five percent to about 20x and it depends on what you're trying to do what type of code how much context it needs a lot of it is related to tuning in the in the system”

0.52

There is no clear, singular inflection point where AI transitions from human-guided improvement to fully autonomous recursive self-improvement; instead, the boundary is fuzzy, and humans will remain drivers of the improvement loop for a very long time.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“I think that it is going to be a much fuzzier boundary for you know getting to self-improvement or or not um I think what will happen is that more and more of the Improvement Loop will be aided by AIS but humans will still be driving it um and it's going to go like that for a long time”

0.50

As foundation models get bigger and capabilities improve, the trend is toward developers achieving more with just prompt engineering and in-context learning (increasing token windows), and this trend grows proportionally to model capability rather than model size.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“as the models get better and better uh it does seem like there is a trend towards less and less of a need to fine-tune and you can do more and more in the context and when you say fine tune you mean changing parameter weights yeah yeah I mean is there going to be ability and ability at all to change the perimeter weights uh in the um the GPT world we yeah we'll definitely offer something there but it like right now it looks like maybe that will be less used than ability to offer like super cheap context like 1 million if we can ever figure that out”

0.50

Parameter count is not the right metric to focus on for advancing AI capability; the gigahertz race in chips during the 1990s-2000s is an instructive analogy—manufacturers pointed to large numbers (gigahertz) that became irrelevant to user experience, so AI should focus on capability and safety rather than parameter scaling.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Sam Altman

“I think it's way too much focus on Perimeter I mean parameter count will Trend up for sure uh but this reminds me a lot of the gigahertz race in chips in the like 90s and 2000s where everybody was trying to like point to a big number and then event like you don't need probably most of you don't know how many gigahertz are on your iPhone but it's fast yeah like what we actually care about is capability”

0.49

Foundation models like GPT-4 are becoming so large (up to 1 trillion parameters) that fine-tuning (retraining model parameters on custom data) is becoming less necessary, and instead developers prefer to use the base model with in-context learning (providing relevant examples and instructions within the token window).

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Sam Altman

“we thought it was going to be much more of a fine-tuning story and we had been thinking about how to offer that in different ways but people are doing pretty amazing things with the base model and for a bunch of reasons often seem to prefer that so we're like actively reconsidering what customization to prioritize given what users seem to want seem to be making work”

0.49

OpenAI is exploring the possibility of 1 million token context windows for GPT-4, which would further reduce the need for fine-tuning customization.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Sam Altman

“we'll definitely offer something there but it like right now it looks like maybe that will be less used than ability to offer like super cheap context like 1 million if we can ever figure that out”

0.46

Altman hopes this moment will be like a chapter in history books, but believes that over the next several billion years, far more significant and exciting things will happen, making this current moment just one small part of much larger progress.

forecasthigh valuespeaker onlynovelty 1/4durability 4/4· Sam Altman

“I mean I I think I hope this will be like uh you know a page or a chapter in history books but I think that over the you know next several billion years uh such unbelievable things are going to happen that this will be just sort of like you know one small part and there'll be new and bigger and more exciting opportunities and challenges in front of us”

0.45

Moving with caution and increasing rigor for safety issues is important as AI capabilities become more serious, but the open letter is not the optimal mechanism to address this problem.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Sam Altman

“so I think moving with caution and an increasing rigor for safety issues is really important the letter I don't think is the optimal way to address it”

0.45

The distinction between what's truly novel and what's been long anticipated is an important question about the timeline for self-improving AI systems.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“there's two questions in there first within open AI how much of a force multiplier do you already see within the creation of the next iteration of the code and then the follow-on question is okay what does it look like a few months from now a year from now two years from now are we getting close to that day where the thing is so rapidly self-improving that it hits some”

0.45

GPT-4 has been described by many as not only the most capable model OpenAI has released but also by far the safest and most aligned, and resists doing harmful things unless deliberately prompted to do so.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Sam Altman

“it's been really nice since we have launched gbt4 how many people have said like wow this is not only the most capable model opening eyes put out but like by far the safest and most aligned and unless I'm trying to get it to do something bad it won't”

0.45

The open letter calling for a six-month pause on AI development lacks technical nuance about where safety pauses are actually needed, and earlier versions incorrectly claimed OpenAI is already training GPT-5, when in fact OpenAI is not currently training GPT-5.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Sam Altman

“I also agree that as safety Cape as as capabilities get more and more serious that the safety bar has got to increase um but unfortunately I think the letter is missing like most technical nuance about what's where we need the pause like it's actually like open AI an earlier version of the letter claimed that open AI is trained in gp5 right now we are not alone for some time um so in that sense it was sort of silly”

0.45

OpenAI does not publicly share the specific cost of training foundation models like GPT-4 beyond noting that it is in the order of tens of millions to hundreds of millions of dollars, and that these costs are rising over time.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Sam Altman

“and then the investment in the creation of the foundation models um is in the on the order of 50 million 100 million just in the in the training process um so it seems like is it now What's the magnitude there we don't share base much more than that okay and and Rising I assume over time yeah”

0.44

The questioner partnered with a VC fund member who wrote an entire book on GPT-2, published on Amazon under a title like 'Start Here' or 'Start Here Romance,' that sold approximately 10 copies (of which the questioner bought one), demonstrating how outdated GPT-2 knowledge is today.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“one of our partners in our Venture fund wrote an entire book on gpt2 and you can buy it on Amazon it's called start here or start here romance I think about 10 copies have sold I bought one of them so maybe nine copies have sold but if you read the book it's just not a good book and here we are it's only that was four years ago it's only been four years and now the quality of the book has gone from you know gbt234 not a good book you know somewhat reasonable book to now it's possible to write a truly excellent book”

0.42

Altman knew ChatGPT would catch fire as a consumer product approximately 48 hours after its launch.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Sam Altman

“in terms of when we knew that Chachi BT in particular was gonna like Catch Fire as a consumer product probably like 48 hours after launch”

0.39

The questioner built neural networks from sub-1-million parameter models, watched them scale to billions (GPT-2 ~1.5B), then to 175 billion (GPT-3), and observed mind-blowing differences in capability at each order of magnitude increase.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“I've been building neural networks myself since the the parameter count was sub 1 million and they're actually very useful for a bunch of commercial applications and then kind of watch them tip into the billion and then the you know with gpt2 I think about one and a half billion or so uh and then gpt3 and now gpt4 so you go up we don't know the current parameter count but I think it was 175 billion um in gpd3 and it was just mind-blowingly different from gpt2 and then gpt4 is even more mind-blowingly different”

0.22

The conversation took place in a room of approximately 900 entrepreneurs and company builders in the ecosystem triggered by ChatGPT, assembled with the goal of helping them build successful companies on the OpenAI platform.

factualspeaker onlynovelty 0/4durability 2/4· Unidentified Speaker — Breakthrough potential of AI | Sam Altman | MIT 2023 [T5cPoNwO7II]

“come on contacts for you I know you're on this worldwide tour uh trying to help control the fire of the gpt4 and Chad GPT have started this particular room is maybe a little different from from a lot of that most of the people here in this room are either building companies or are working on plans to build companies that are in the ecosystem really triggered by chat gbet”

0.19

The model context window is currently 32,000 tokens for the largest GPT-4 variant and 8,000 tokens for the base variant, according to statements in early 2023.

factualestablishednovelty 0/4durability 1/4· Sam Altman

“I think you have 50 000 tokens to play with right now in the basic model is that right about 32 000 in the biggest model thirty two thousand eight thousand base one okay”

0.17

Altman does not claim to be an exceptional business strategist despite Paul Graham's reported characterization, and would defer to people with actual practice in building fast-growing consumer or enterprise companies on platforms like iOS.

factualspeaker onlynovelty 0/4durability 2/4· Sam Altman

“I don't think PG quite said that I know you said something like in that direction but uh in any sense in any case I I don't think it'd be true uh I think there are people who are like unbelievable business strategists and I'm not one of them so I I hesitate to give advice here”