"Where is all the automation?" — Diogo Almeida, TypeSafe AI — Transcript
Full transcript
- 0:01[music]
- 0:12>> I'm Diego Almeida, and this talk is I
- 0:15Love You So Jev Much. Um,
- 0:18Glass Animals reference. We're not just
- 0:20weird. Um,
- 0:22like I said, I'm Diego Almeida. I'm
- 0:24known as the Jev guy, sometimes Jev pink
- 0:27man. Um, also co-created ChatGPT, RLHF,
- 0:31and InstructGPT, but I assume that for
- 0:33the crowd today, y'all know who I am, at
- 0:36least mostly.
- 0:37[snorts] Um,
- 0:38I have like a bunch of AI-specific stuff
- 0:41I want to get to, but I like going into
- 0:42my favorite topics in AI. First one is
- 0:46what the [ __ ] is actually going on and
- 0:48where is all the automation? If you've
- 0:50heard me say this before, I'm sorry, but
- 0:52it's such an important thing to beat
- 0:54into the field cuz people still do not
- 0:56get it, and I need to yell this until
- 0:58everyone really understands what the
- 1:00[ __ ] is going on. And if you're in AI,
- 1:02you should have a really good answer
- 1:03about why AI can do so many awesome
- 1:06things and yet still be so useless for a
- 1:08bunch of other ones.
- 1:10My relationship to ChatGPT is somewhat
- 1:12unique, one of the few people who hates
- 1:14on ChatGPT, who made it. I don't hate
- 1:18ChatGPT, to be clear. I love using
- 1:20agents. I love using coding agents. I
- 1:22love chatbots. They're just quite
- 1:24limited. And in particular, I think it's
- 1:26a little tragic how much they've
- 1:28bifurcated the field into overpromise,
- 1:30underdeliver.
- 1:31Really, really terrible stuff for the
- 1:32world in my opinion.
- 1:34And I like to talk about like the two
- 1:36cults in AI and how we can resolve the
- 1:39cognitive dissonance that everyone I
- 1:41think should be having unless they have
- 1:43a very good answer.
- 1:45There's cult one, which is you know, the
- 1:46cult of like EAs. They look at all the
- 1:49benchmarks and they say like not only
- 1:50are LMs surpassing human level on
- 1:53basically every benchmark, they have
- 1:55been for two years. We've been
- 1:56announcing AGI three months away for two
- 1:59years now.
- 2:00Even since like the '03 days. Do people
- 2:01remember that? No one even care-cares
- 2:03about that model anymore and it it was
- 2:05that that was like peak internal to
- 2:07OpenAI AGI hype.
- 2:08And on the other hand, you look at the
- 2:10non-benchmarks and you look at the
- 2:12reality and you see the rest of the
- 2:14world, especially if you guys go outside
- 2:16and leave San Francisco, a lot of the
- 2:17world is like really unchanged despite
- 2:20AI being so [ __ ] cool. And I think AI
- 2:23is transformative. It's it's a rare
- 2:24point of view to actually think AI is
- 2:26cool and transformative and also like
- 2:29underwhelming the world right now.
- 2:32And the tricky part is this is the same
- 2:33field and we are all seeing the same
- 2:35facts. So what gives? You know, why is
- 2:38AI jagged, so to speak? Like why can it
- 2:40be so good at some things and so not
- 2:43good at other things? And what is the
- 2:45Occam's razor explanation for that? Um I
- 2:48will be blasting through my slides fast.
- 2:50I'm sorry. Maybe I could slow down for
- 2:52people taking pictures. Um but uh I I'm
- 2:55noticing that uh
- 2:57this is going to go fast is all I'm
- 2:58saying. So I think that we as a field
- 3:01should have an Occam's razor explanation
- 3:03for why are there some things that are
- 3:05just
- 3:06not just good, they are unbelievably
- 3:08good, superhuman, like millennium prize
- 3:10problems in math. But on the other hand,
- 3:12they're not just bad, they're so
- 3:14useless, they can't even be used to
- 3:15touch accounting. That sounds way easier
- 3:17than million prize problems in math,
- 3:18right? I could do the latter. I
- 3:20definitely could not do the former and I
- 3:21was a math athlete. So what gives here?
- 3:24What is the reason for this? And I
- 3:26normally love this as an interactive
- 3:28thing, but I will instead just tell you
- 3:29my answer. Sorry for the power dynamic.
- 3:31Um everything on the left side is
- 3:34intrinsically a human in the loop task.
- 3:36Like deep research, I could not do
- 3:38better than deep research. The models
- 3:39are so [ __ ] good at that.
- 3:41But if gun to my head or gun to your
- 3:44head, you had to pick between me and
- 3:45deep research and the answer had to be
- 3:47correct, not well written, you'd
- 3:49probably pick me every time. And that's
- 3:51the distinction between like optimizing
- 3:54for automation, which tends to be a lot
- 3:56easier and a lot simpler, yet
- 3:58AI keeps failing at this. Every year
- 4:00there's like an infamous failure of AI
- 4:02failing to even automate drive-throughs.
- 4:04And that should be kind of shocking to
- 4:06people. And my claim is actually for
- 4:09these tasks pleasing a human is nothing
- 4:11to do with doing the task. What you want
- 4:13to do is basically be an abstraction
- 4:16away. You want the decision to be made
- 4:17and then you want to get back into doing
- 4:19automation and code is the language of
- 4:21automation. Therefore you need to
- 4:22integrate with code. This is a lot but I
- 4:24will summarize it again. I get the irony
- 4:27that uh you know I co-created RLHF and I
- 4:30do believe RLHF is part of the problem
- 4:32and RLHF in particular is why all AI
- 4:35today does require a human in the loop
- 4:38with an asterisk unless the task is like
- 4:40very trivially verifiable and can run on
- 4:42its own but still requires a human in
- 4:44the loop to verify it afterwards. It is
- 4:46because we literally put it in. That is
- 4:48what the HF is in RLHF and my claim is
- 4:51that there's a better way. You know my
- 4:53summary of the state is 100% of LMs
- 4:56today are optimized for assistance. You
- 4:58might talk about RLVR but the RLVR
- 5:00models are very bad to interact with if
- 5:02they've not been RLHFed.
- 5:03The result is LMs become too good to be
- 5:06true at assistance. They are just so
- 5:08[ __ ] good like you should never pick
- 5:09me over ChatGPT for your ChatGPT
- 5:11queries. But for actually doing like
- 5:15you know run in the background cognitive
- 5:17work you should not trust LMs at all
- 5:19because they've not been optimized for
- 5:20that which is why they're so bad at the
- 5:22tasks of automation and I would also
- 5:24guess that automation is a lot more
- 5:26valuable than assistance because you can
- 5:27do way cooler deeper I'm going to get
- 5:29into this later but I think there's
- 5:30going to be so much stuff that we can do
- 5:32and we are just scratching the surface
- 5:34of that right now and that's the tragic
- 5:35thing about AI uh the state of AI for
- 5:38me.
- 5:39Let's talk about the Jevolution. Um
- 5:43I I can't tell if this is appropriate
- 5:44but it's too late.
- 5:45Um
- 5:46oh god don't take pictures of this one.
- 5:48Um
- 5:50cool.
- 5:51Oh I'm getting a thumbs up. Yeah
- 5:53excellent. So we're we instead of doing
- 5:55RLHF we just have to do something else.
- 5:57We need to think from first principles.
- 5:58How do we optimize for machine native
- 6:00AI? You guys hopefully know the
- 6:01interface of Jev and that is what we are
- 6:03doing. More important than optimizing
- 6:05for its interface, we are
- 6:07uh rethinking how AI should be done
- 6:10because this is what we did again for
- 6:11RLHF.
- 6:13For Sutton's bitter lesson,
- 6:15I'm going to loosely paraphrase to make
- 6:16my point easier to say, is that
- 6:19scaling um and compute matters more than
- 6:22algorithms. This is why people become
- 6:23like scaling pill geronimous and I would
- 6:26argue that there's an even bitterer
- 6:28lesson than that, which is data always
- 6:30matters more than compute. We are
- 6:31learning that quite a bit in the field
- 6:33now, but it has been obvious since the
- 6:35beginning of time. I'll tell you when
- 6:37the this is the one you should be
- 6:38photographing. I would argue that the
- 6:40the bitterest lesson is doing the right
- 6:43task. And doing the right task is like a
- 6:45weird thing to think about, especially
- 6:46for ML people, cuz they like algorithms.
- 6:49They want number goes down or up or
- 6:51whichever is good direction. They don't
- 6:52want to think about the real world. But
- 6:54this is always how you make gigantic
- 6:56gains in performance. This is how the
- 6:58big breakthroughs happen. 100% of the
- 7:00time it comes from this. And this is
- 7:02true for classical RL games. This is why
- 7:04Sutton had his bitter lesson, it which
- 7:06is data is derived by compute when
- 7:09you're doing rollouts. Like there's
- 7:10there's why would you need off policy
- 7:12data when playing chess or something
- 7:13like that? And doing the right task is
- 7:15defined by the rules. So actually when
- 7:17you're doing RL, you don't even need to
- 7:19think about the right task.
- 7:20For pre-training, we actually as a field
- 7:23had to do a lot of thinking about the
- 7:24right task. I don't know if people
- 7:26remember the Bert days with um masked
- 7:28language modeling, but it took a while
- 7:31and it was very unclear to figure out
- 7:32that next pro token prediction was super
- 7:35[ __ ] cool, not because it was better
- 7:37at masked language modeling. You know,
- 7:38the objectives on their own aren't
- 7:40special. It's about what you can do with
- 7:42them afterwards. And it turns out next
- 7:44token prediction is incredible at
- 7:45generating strings, which is what these
- 7:47pre-trained models were so good at. So
- 7:49no level of masked language modeling,
- 7:50even if you had today's compute and
- 7:52algorithms, would make it as interesting
- 7:54as GPT-3 was because it was just doing a
- 7:56different task. And data, you know, the
- 7:59internet is obviously a giant
- 8:00bottleneck. Now we're like, you know,
- 8:01burning books or whatever in order to
- 8:03get more data, but data obviously
- 8:05matters a lot more than compute and
- 8:06people keep like rediscovering this and
- 8:08figuring this out.
- 8:09It's extra true for making LLMs useful.
- 8:12Um, before RLHF even existed, the data
- 8:15for RLHF didn't exist and you could
- 8:18uh like it was night and day. I'll just
- 8:19show you the next slide. This is the
- 8:20probably the relevant part. This is what
- 8:22this was the work in the RLHF paper.
- 8:25And the high level was that even with
- 8:27the dumbest algorithms in the world,
- 8:28even with a GPT-2 sized model, you got
- 8:31something like GPT-9 level performance,
- 8:34like real GPT-9, not today's like fake
- 8:36GPT numbers. Like the numbers used to
- 8:38mean like a 100x scaling in effective
- 8:40compute for pre-training and they've
- 8:42given up on that because it just can't
- 8:43scale anymore. And still even with that
- 8:45it would take something like GPT-7 to 9
- 8:48to match GPT-2, the model from like 2018
- 8:51or whatever it was, at the task of
- 8:53instruction following. That is kind of
- 8:55wild because you think big LLMs are so
- 8:58important for this and they are, but
- 9:00that is not the most important thing
- 9:01compared to even the gigantic jump of
- 9:04doing the right task. And spoilers, this
- 9:06is what we did for Jev, right? Like we
- 9:08made this gigantic leap in the field
- 9:09that no one could realized could be
- 9:11coming because we were optimizing for a
- 9:13task and very few people think about the
- 9:15right task.
- 9:16Um,
- 9:17this is why Jev has the je ne sais quoi.
- 9:21Um, that's a really smart pun I think.
- 9:24Um,
- 9:25that that little bit of a pity laugh,
- 9:27but thank you. Um, you get what you
- 9:29optimize for. I'm looking for a better
- 9:32word than reliable, so if you guys have
- 9:33suggestions let me know. But there's if
- 9:35there's this reliability of intelligence
- 9:37that's really missing in LLMs. I often
- 9:39call it you get Einstein 95% of the
- 9:42time, but Mr. Bean 5% of the time.
- 9:45And if you can't tell which is which,
- 9:46how the [ __ ] are you going to automate
- 9:47anything? And instead instead of like
- 9:49focusing on like sometimes occasional
- 9:51smarts that you can then verify, what if
- 9:53you focused on extreme extreme narrow
- 9:56focused reliability so you can get that
- 9:58smarts every time? And this doesn't show
- 10:00up on benchmarks. It is the reason why
- 10:02people love Jev. And this is also I
- 10:04realize in hindsight we obviously didn't
- 10:05prepare for a chat GPT moment for
- 10:07developers, but it happened because
- 10:09people in the AI space just overpromise,
- 10:12underdeliver so [ __ ] much that having
- 10:15a model that actually didn't do that was
- 10:16a surprise. And it is not the speed,
- 10:18cost, or even interface. This is
- 10:20something that, you know, my team tells
- 10:22me I shouldn't be telling people so that
- 10:24we can have a long-term advantage, and I
- 10:25just care too much. Like it's a reliable
- 10:27intelligence. Make sick models, please,
- 10:30even if we, you know, if it doesn't
- 10:31involve our models.
- 10:33Um we see that Jev is like a foundation
- 10:36for smart software. I'm going to skip
- 10:37all over this, but I think software is
- 10:38going to be so [ __ ] back. I think the
- 10:41narrative of SaaS apocalypse is
- 10:42obviously not going to happen, but
- 10:45I think that
- 10:46even calling it a SaaS apocalpyse is
- 10:48like an understatement, and I'll get
- 10:49into that a bit, and you guys ideally
- 10:51are the perfect crowd for this. I'm I'm
- 10:52so excited for this part because that's
- 10:53the part that's new to me. Um
- 10:56I want to make a distinction between
- 10:58smart software and just-in-time
- 11:00software. The the crazy part is AI is so
- 11:03transformative, and yet somehow despite
- 11:06AI being so awesome, software has
- 11:08essentially unchanged in the last decade
- 11:10except being cheaper to write. Isn't
- 11:13that what weird? Isn't the it weird that
- 11:15the best we can do in software is having
- 11:16a chatbot to the side that sometimes
- 11:19interacts with the software but cannot
- 11:20take expensive decisions for you? That
- 11:22is kind of like a weird thing if AI was
- 11:24really that transformative, and I think
- 11:26the problem is we're focusing too much
- 11:28on coding agents and not enough about
- 11:30putting the intelligence into the inner
- 11:32loop, which I'm going to get more into a
- 11:33bit. And people are jeving. There's been
- 11:36tons of successes. We don't need to flex
- 11:38this. There's so many cool things, but
- 11:40my AI E secret is going to be that the
- 11:43field isn't jeving hard enough. And this
- 11:45is what Swix said that I can make a call
- 11:47to action for you guys that's unique,
- 11:48and hopefully I have enough time for
- 11:49this.
- 11:50I think the world still has LLM
- 11:52Stockholm syndrome. We are still
- 11:54thinking from a really LLM-filled place,
- 11:56which is
- 11:57I think tragic, you know? Like I I I
- 11:59love the software saying the tools you
- 12:01use change the way you think, and I
- 12:03think that it's changed the way we think
- 12:05way too much,
- 12:06and we're not really rethinking software
- 12:08like I think the field needs. I think
- 12:10that we could just just doing so much
- 12:11more. Hopefully I can try to do my best
- 12:13for my remaining time to convince you
- 12:15guys of that.
- 12:16Um, so let's jump to the future.
- 12:19Um,
- 12:20I an analogy I love, bear with me, is
- 12:23about LEDs. Before there was a blue LED,
- 12:26there was red and green LEDs. These were
- 12:28really limited in capability, and they
- 12:30were used for like simple indicators and
- 12:32you know like clocks and stuff like
- 12:34that. And when the blue LED was made
- 12:37discovered, I actually don't really know
- 12:38how things outside of computer science
- 12:40work. Um, but when it existed, I guess,
- 12:43um, it enabled not just like 50% more
- 12:46use cases, which what you'd expect about
- 12:48like changing colors from two to three,
- 12:50it enabled so many more. Like now LEDs
- 12:52could be used for white light. It could
- 12:53be used for like the displays that we
- 12:55all use right now, and it created a
- 12:57revolution in lighting. And it's it's
- 12:59it's unfathomable how much bigger
- 13:02the capabilities of LEDs were just you
- 13:04know adding a new primitive that
- 13:06composes together with the old
- 13:08primitives. You can see where I'm going
- 13:09here maybe. And I think it's a similar
- 13:12story for software. Software is so
- 13:15[ __ ] incredible. Ignore AI for the
- 13:16last you know half decade, but
- 13:18software's done so many insanely cool
- 13:21things. We've done a lot of stuff you
- 13:23know like do I
- 13:25love using SAS? Not really, but does it
- 13:27solve problems? Abso [ __ ] lutely. And
- 13:30it's incredible how much we've done with
- 13:32three [clears throat] fundamental logic
- 13:33gates. You know like we are doing this
- 13:34with like simple binary operations that
- 13:37can do things like greater than, less
- 13:38than, equal. It's kind of nuts how much
- 13:41software has been built from such simple
- 13:43primitives. And my claim is that what we
- 13:47are doing with AI now and we're making
- 13:49it our intelligence machine native, this
- 13:51is kind of conceptually adding a fourth
- 13:53semantic logic gate.
- 13:55What if we can program with common
- 13:57sense, with intention, with human
- 13:59understanding? Like do we still need
- 14:01regexes with like crazy things or can we
- 14:03just say like we want a Gmail email
- 14:04address or a real looking one? There's
- 14:06so much that we can do that we are going
- 14:08to need to push the you know push the
- 14:09frontiers of cheap, smart, fast in order
- 14:12to get done. But I think there's going
- 14:14to be like so many things here, And I
- 14:16think just like the blue LED, it's going
- 14:17to be way bigger than just increasing
- 14:20the representational ability of software
- 14:22by 33 % cuz it's adding effectively
- 14:25another kind of sort of logic gate. And
- 14:27this sci-fi is the stuff that I think
- 14:29will be impossible to predict. Just like
- 14:31who could predict how powerful the
- 14:33internet would have been because it
- 14:34requires it's like much more grassroots
- 14:36and decentralized and it requires people
- 14:38to be empowered to do insane stuff.
- 14:40Um I call this software research. Um
- 14:44that's not a joke. Um but uh I I think
- 14:47that there's going to be like an
- 14:49explosion of what software can do and we
- 14:52will want people to be exploring those
- 14:54boundaries of software, not necessarily
- 14:56because it's valuable, but because they
- 14:58can, because we have like new frontiers
- 15:00to explore. And this is what the early
- 15:02internet was like. Should all of those
- 15:04websites existed?
- 15:05Probably not, but like it was very cool
- 15:08and we found a lot of awesome stuff from
- 15:09that and I think we we you know we have
- 15:11another opportunity for that. So we
- 15:13should not just keep living in this
- 15:15little you know fish tank that we are
- 15:17the victims of open AI and Anthropic on
- 15:18whatever they ship and we just have to
- 15:20be like LLM rappers. So I'm going to
- 15:22pitch software research as a direction
- 15:24here. The theme is going to be real
- 15:27software engineering, not just vibe
- 15:29coding. There's it's going to be like
- 15:31the kind of stuff that everyone has
- 15:32known if you've been a software
- 15:33engineer. Decomposition, breaking tasks
- 15:36down. We need to think about stages,
- 15:37ideally ones that can be verified so we
- 15:39have real encapsulation. And we all need
- 15:42to actually do state management instead
- 15:44of assuming everything is global state
- 15:46within a KV cache. And the theme is just
- 15:47going to be like real [ __ ]
- 15:49engineering by people who care a lot
- 15:51about the end result. It doesn't have to
- 15:53be valuable. Ideally it's sick as [ __ ]
- 15:55Um and you know we should be pushing the
- 15:57limits.
- 15:58I'm going to go into a couple of like my
- 16:00favorite ideas and I want to figure out
- 16:02how to encourage these in the world. So
- 16:04this is my first step at this. I'm just
- 16:06going to talk about it. We have raised a
- 16:08bunch of money so I would like to figure
- 16:10out how to fund this, but we don't have
- 16:11the operations help to do that. But here
- 16:13are just some ideas out there. Um I
- 16:15think agents are
- 16:18they're cool, but like they're
- 16:20they're so inflexible in their design
- 16:22because of the tyranny of the KV cache.
- 16:24I don't have the time for this if this
- 16:25timer is to be believed, but believe me
- 16:28that there's so much that people have
- 16:29tried to do but failed at because it
- 16:31ends up becoming inefficient because it
- 16:33fights against the KV cache. But I think
- 16:35there's a lot more to do. Why is memory
- 16:37O of N, right? Like why can't we have
- 16:39all of our history of coding agents
- 16:41always included in every new call? That
- 16:44is how continual learning works, right?
- 16:47And we have the memory available. It's
- 16:48just that we don't have the smarts to
- 16:50map reduce over the memory to find out
- 16:52like what you should what you should
- 16:54remember, what should you shouldn't. Um
- 16:56I think there's many ways to do login or
- 16:57even better lookups. Why when you're
- 16:59making many agents at once, why can't
- 17:01they not share state? Why do they need
- 17:03to rediscover the same patterns of the
- 17:05codebase over and over and over again?
- 17:07And why can't non-naive routing in sub
- 17:09agents work? Like even naive routing
- 17:10doesn't work. But like why can't we do
- 17:12these things? And why just like the you
- 17:14know the the big labs keep going back
- 17:15and forth on whether or not sub agents
- 17:16work? Um
- 17:18I think that like from an engineering
- 17:20perspective, there's elegance to it, but
- 17:22they fight against the KV cache. And I
- 17:23think that if we design agents around
- 17:25that, we could do so much more.
- 17:28This one's the probably the most vague
- 17:30of my directions, but I feel very
- 17:32strongly about it. Um the world is full
- 17:34of principal agent problems. This is
- 17:36when you have like you know ulterior
- 17:37motives between this is not a coding
- 17:39agent, but like any agent. The the
- 17:40classic example is citizens versus
- 17:42elected officials. And I don't want to
- 17:44make a too political claim here, but
- 17:46like there's misalignment in all levels
- 17:49of society. And this results in like
- 17:51waste, fraud, bad things, and it it
- 17:54results in for a society that wants
- 17:56fairness, you need to over emphasize the
- 17:58letter over the spirit, which then
- 18:00results in loopholes and pay pay to win
- 18:03dynamics in the world. And I think the
- 18:05root of that problem is we can't encode
- 18:06the spirit. Just like with software,
- 18:08we're constrained to encoding bits
- 18:10instead of what we actually meant in
- 18:12that situation. And I think so many
- 18:14organizations, society could be
- 18:16transformed by really encoding the
- 18:17spirit all over the place.
- 18:19Obviously real-time intelligence
- 18:21security,
- 18:22can we prevent the OpenAI hugging face
- 18:24sandbox, blah blah blah? Um, why did why
- 18:27couldn't they prevent it? It is not
- 18:29because they didn't have the
- 18:30intelligence. I would argue it's a lot
- 18:31easier to figure out they're doing a
- 18:32naughty thing than to actually do the
- 18:34naughty thing. It is because it was not
- 18:36fast, cheap, and reliable enough to
- 18:38actually run real-time. And what happens
- 18:41when every corporation has their own
- 18:44real-time security that's actually
- 18:45smart? How cool would that [ __ ] be?
- 18:48I can't talk about generative UIs
- 18:50because people are going to be talking
- 18:51about it afterwards. I don't want to
- 18:52steal their thunder, but it's cool as
- 18:54[ __ ] trust me. Um,
- 18:56like semantic dependent typing. Why
- 18:59can't
- 19:00the we have like a type system that
- 19:02actually is semantic, right? Like why
- 19:04can't we
- 19:05like declare what local properties we
- 19:08want so we don't have to read a chain of
- 19:10dependencies deep down?
- 19:11These all should be possible. And
- 19:13imagine if you could like run a JEP-like
- 19:15model in your lint checker to check if
- 19:17your it's all type-safe and it could you
- 19:20know uphold semantic properties. It
- 19:21would be so [ __ ] cool.
- 19:24There's there was a great Twitter video
- 19:26on um like a voice interface. In
- 19:29general, I think new interfaces are so
- 19:31exciting. This is to me the return of
- 19:33like this is going to show by agent
- 19:34software like do what I mean. This used
- 19:36to be like a popular thing, but it never
- 19:38really worked because you it was it was
- 19:40too many hard-coded rules to really
- 19:41figure out what I mean. Um, the stopping
- 19:45slop obviously a big theme of here, but
- 19:47like we definitely believe this. This is
- 19:49my props I have to
- 19:50We have a stop Gen AI baseball bat for
- 19:53type-safe. It is very very awesome. I
- 19:56want to talk about the root of slop. Um,
- 19:58the the root of slop is that AI is
- 20:00unreliable. Therefore, you as a company
- 20:02or you as you know like someone who's
- 20:04doing stuff, it can't be used for
- 20:05decisions that costs you stuff.
- 20:08But it's actually totally fine to push
- 20:09the costs into other people, which is
- 20:11horrible, you know externalized costs.
- 20:13And then this is what ends up becoming
- 20:14slop.
- 20:15You know if people actually read this
- 20:16stuff and filter it and like care about
- 20:18the quality, it becomes not slop. But
- 20:20now the world is full of this because
- 20:22there's an asymmetric advantage for
- 20:24people who are like bad actors and don't
- 20:25give a [ __ ] So, I think that we can
- 20:27save the world from at least inbound
- 20:29hell among many other things and that
- 20:31would be very cool. Um, and even
- 20:34something I think is super sick is
- 20:36cursed software. Like why? If you guys
- 20:38ever seen the subreddit bad UI battles,
- 20:41it is so [ __ ] cool even though they
- 20:43make the horrible interfaces like this
- 20:45whole game of like would you like to
- 20:46delete your account becomes, you know, a
- 20:48a random thing. And I actually think
- 20:51doing things because we can is what
- 20:53software engineering is about cuz who
- 20:55knows maybe 1% of these ideas actually
- 20:57lead to something super awesome. And I
- 21:00think like really cultivating even bad
- 21:02ideas is kind of what being like this,
- 21:04you know, playful hacker is all about.
- 21:07And generally just automating
- 21:09everything, this is what the follow-up
- 21:11this is this is a flop image but uh so
- 21:13our logo's a little bit wrong but don't
- 21:15stare at that that's a little weird. Um,
- 21:18this is what a lot of the follow-up
- 21:19talks are going to be about. Thank you.
- 21:21Um,
- 21:22and and if you're a software engineer,
- 21:24you almost certainly think in the world
- 21:26of bits. If you know anything about
- 21:28software uh and you've interacted with
- 21:30software, you know that there are things
- 21:33software can do and software cannot do
- 21:35because that is just only obvious,
- 21:37right? Like there's some things that you
- 21:39know you need to do because that is not
- 21:41compatible with software. But what if
- 21:43those constraints were no longer true?
- 21:45Maybe not today but maybe in like 6
- 21:47months or a year, what couldn't we
- 21:49automate? And that's kind of where I
- 21:50want to leave this whole thing because I
- 21:52think there is very likely to be a
- 21:54future where it's not a question of
- 21:56whether or not we can automate
- 21:57something, it's going to be a question
- 21:58of like how much work is it which is how
- 22:00it is right now for the stuff that can
- 22:02be automated and that would be cool as
- 22:04[ __ ]
- 22:05So, hell yeah. Thank you.
- 22:07Um,
- 22:09I'm
- 22:10super overtime but thank you to the
- 22:12early believers. You know, we had no
- 22:14idea so many people would resonate with
- 22:16this. We thought we were the insane ones
- 22:18but it turns out like we were actually
- 22:20the majority of the world who realized
- 22:21AI could be so much more. Thank you to
- 22:24like the people who are like realistic
- 22:25about reliability. Everyone knows I hate
- 22:28benchmarks and a lot of other people are
- 22:29really screaming out like benchmarks
- 22:31[ __ ] suck. They're actually
- 22:33they hurt the people who do not have
- 22:35time for anything other than benchmarks
- 22:36most. It's tragic. And for the people
- 22:38who are pushing the limits of software,
- 22:40like [ __ ] yes. And we want to be there
- 22:42with you helping you all do that. The
- 22:44the dreamers and the people who don't
- 22:46want to be limited by everything that
- 22:48has happened so far, we want to really
- 22:51fight for you and we want this to be as
- 22:54reliable as a fourth logic gate. And
- 22:56that is a very high bar and we have like
- 22:58a lot of work to do. So thank you, thank
- 23:01you.
About this transcript
This page contains the full transcript of "Where is all the automation?" — Diogo Almeida, TypeSafe AI by AI Engineer, generated from the public captions YouTube serves with the video. The transcript has 4,767 words across 692 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.