# ai make ai
every person had to learn a bitter lesson or two in their lives
maybe their first love rejected them, maybe they learned they're not invincible the hard way
normal people learn it in different ways, but machine learning engineers seem to only have one, and it is the way of manually crafting algorithms and datasets, only to find out that models just want to learn, and applying more compute to a problem often is the best way to solve it in the long run
the recent turing award laureate richard sutton popularized (introduced?) this idea in his famous essay:
>> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great.
>>
>> The two methods that seem to scale arbitrarily in this way are search and learning.
read [the whole thing](http://www.incompleteideas.net/IncIdeas/BitterLesson.html), it's under 2 pages
**
and yet, we write prompts by hand and we build ai agents using a dozen of frameworks, of which i tested a bunch:
![[Pasted image 20250331233546.png]]
[autogpt](https://github.com/Significant-Gravitas/AutoGPT) was the first framework of this kind, released in march 2023, babyagi, superagi, gpt-engineer and dozens more followed shortly
more recently, releases like deep (re)search by google, openai, perplexity, and grok can be seen as agentic, as well as anthropic computer use, openai operator, and manus - agents that can click around in the browser or the operating system. i describe this capability as [[screen navigation]]
autogpt & co have never worked - rather, they failed completely and miserably, casting doubt that the idea of agents makes any practical sense
---
## the year of agents
recently though, things changed
the aforementioned deep research tools are now capable of saving you hours and days of work for software development, and agents such as [[knowledge/tools/bolt.new]] and [[cursor]] can build entire apps with just one prompt
interestingly, one of those agentic tools, lovable, started as gpt-engineer - i tried running thrice at different points in time, it always failed but lovable works decent (even though it makes a landing page no matter what you ask it to do)
let's have a look at 60 sec demo that is both boring af and should send shivers down your spine if you fancy extrapolating trends into the future
>> design and build a website card for a portfolio of an architect
![[2025-03-31 23-55-19.mkv]]
the apps coming from those coding agents are quite simple but a lot of the time, they run on the first try, which wasn't the case just half a year ago. some of this progress is due to work on bettering the agentic architecture, some of it is due to the release of reasoning models. although both bolt.new and cursor work quite well on sonnet35, which has been released in june 2024 - feels like ages now!
**
when companies like anthropic and openai talk about 2025 being the year of agents, that's what they mean. more and more companies keep releasing their own versions of researching and coding, and other domains. but of course, as of march 2025, humans still need to write prompts, formulate requirements, go to webpages, install apps, send emails etc - and ultimately, decide what those agents do. humans define intentions of ai, for now
but what if bolt doesn't work for your use case, which are pretty much all but fairly simple frontend-focused use cases, such as landing pages, business card websites, and some dashboards? then you're screwed, because it won't code in python, mostly due to the limitations of the runtime environment. cursor does backend, writes tons of code but its apps rarely run
if you would want to improve the agent (introduce best practices for security), constrain it to your use case or code base (use firebase instead of supa), you wouldn't be able to do that *easily*. deep research, computer use, coding agents are still static pieces of software like any other, defined front to back by their developers. wanna change it - the best you have is to fork [openhands](https://github.com/All-Hands-AI/OpenHands) (it never works btw) and spend 30-50-100k tinkering with it
and certainly, you wouldn't be able to create your own agent by telling ai to build one from scratch
all while a bunch of sites and apps are just the same CRUD app with different design (still built from the same components like material design or whatever) - and that's just bad user experience
which reminds me of 2016, when the first wave of chatbot craziness happened, and companies like google, facebook, microsoft, telegram, kik, slack, ibm, and others have released their own chatbot frameworks. in the time of "sorry i did not understand your request", i tested 2000 of them - all were bad, mostly due to ai models being fairly incapable, and also because not developers had no idea what they were doing. frustration and disappointment partly caused by this testing made me abandon natural language processing for several years - a very stupid decision in the hindsight
one of the things that telegram did right back then, was to create a tool they called [botfather](https://telegram.me/BotFather), a bot that creates other bots
![[Pasted image 20250401001443.png]]
chatbot developers would use to instantiate their own chatbot in the system, give it a @name and registers with the platform, so that the dev can link their code - and users can talk to it
today, there's no way for people to create their own agents. the reason is the complexity and the fragile structure of the agentic systems compared to the bots of yesteryear - hard to build them, hard to evaluate them, hard to maintain. but this will be done, and we already know how
---
## building ai that teaches itself
a group of researchers under [[jeff clune]] recently published a paper titled automated design of agentic systems, exploring the possibilities of ai agents discovering and building other agents
![[Pasted image 20250301231330.png]]
watch "[can ai improve itself](https://www.youtube.com/watch?v=1kwbp8hRRfs)" by machine learning street talk, it is the best ai podcast in existence and one of their deepest and most insightful episodes
gist:
- researchers define [[Meta Agent Search]] that is unbelievably simple
- MAS analyzes the task, creates an agent workflow, code the task in terms of prompts, writes python code, executes, evaluates the results of the task, and if the agent was unsuccessful, it changes the prompt, adopts the code, tries again
- after a bunch of iterations, the meta-search agent happy confirms the explanation and stores it
- no humans involved
- MSA is itself being iteratively improved, ie, it is learning to learn
![[Pasted image 20250401005933.png]]
that's it. their agents achieve SOTA on some benchmarks, and they define a new field of study, automated design of genetic systems. while a bold claim, i think they are generally correct in their approach and ADAS will become the way to build AGI over the next 3-5 years
**
another example comes from a different, although partly the same team
it's called [ai scientist](https://arxiv.org/abs/2408.06292). the purpose of this agent is to do research in artificial intelligence and it works as follows:
- the agent considers a research question
- designs experiments
- writes code for them, runs the experiment, evaluates results, makes charts
- then writes the whole paper based on the results, with introduction, literature review, and conclusions
- then it goes through a simulation of a peer-review process, where it analyzes its own work, and it satisfies scores in the library for the future works to build upon
fun fact:
ai scientist wrote its first peer-reviewed paper two weeks ago, see [tweet](https://x.com/SakanaAILabs/status/1899646987781501181)
<blockquote class="twitter-tweet"><p lang="en" dir="ltr">The AI Scientist Generates its First Peer-Reviewed Scientific Publication<br><br>We're proud to announce that a paper produced by The AI Scientist-v2 passed the peer-review process at a workshop in ICLR, a top AI conference.<br><br>Read more about this experiment → <a href="https://t.co/LpLYLnZMCQ">https://t.co/LpLYLnZMCQ</a>… <a href="https://t.co/z18yJB945y">pic.twitter.com/z18yJB945y</a></p>— Sakana AI (@SakanaAILabs) <a href="https://twitter.com/SakanaAILabs/status/1899646987781501181?ref_src=twsrc%5Etfw">March 12, 2025</a></blockquote>
**
those two papers give us a glimpse into the future
on one hand, we know that ai is capable of coming up with new research questions about itself, run experiments, and most importantly, produce novel insight
on the other hand, we have the meta-search agent that uses evolutionary algorithms to improve prompts and code, and is capable of running a whole family of agents
where does it lead us?
---
## bitter lesson all over again
so now, it is time to consider the bitter lesson again and see what we know
1. search and learning scaled with compute outperform any human crafted approaches
2. human-designed agents are here and perform quite well, will only improve as foundation and reasoning models get better, as well as our agent architectures
3. ADAS and ai scientist give us a proof of concept that agents can effectively create other agents (micro level), and that ai can already work on self-improvement successfully (macro level), thus with (1)
1. --> ai will get better at creating ai than humans with enough compute
oh, then there is the botfather idea
bringing this all together, we can see that
*creating production grade system and an interface to an agent that makes agents will result in a kembrian explosion of human designed agents that can be used as seeds and evolved in a metalearning loop*
in other words
> get a big gpu
> make a botmaker
> tune and metaprompt
> invent AGI
we already know the way there. we have the algorithms, we have neural architectures, we have dirt-cheap apis, and advanced metalearning techniques. ai that build ai is coming, and you better be ready for it - making an agent won't cost 100k but 100
for now, i am reading everything that jeff clune, his phds, and co-authors publish
to start, run the following prompt with whatever deep research you prefer:
```instruction
Title: Exploration of Open-Endedness, Meta-Learning, Prompt Optimization, and Evolutionary Algorithms
Objective:
• Provide a comprehensive analysis of how open-endedness, meta-learning, prompt optimization, and evolutionary algorithms interrelate.
• Discuss contemporary applications and potential future research directions.
• Illustrate how these concepts might be combined in advanced AI research.
Instructions:
1. Introduction:
• Define each concept (open-endedness, meta-learning, prompt optimization, evolutionary algorithms).
• Briefly explain how these concepts are relevant to advancing AI.
2. Open-Endedness and Evolutionary Algorithms:
• Clarify the meaning of open-endedness in evolutionary systems.
• Highlight notable examples or studies that demonstrate open-ended innovation or generation.
• Examine how evolutionary algorithms facilitate exploration and adaptation.
3. Meta-Learning Approaches:
• Describe meta-learning and its significance (e.g., learning to learn, adapting to new tasks quickly).
• Identify how meta-learning algorithms incorporate or benefit from evolutionary techniques.
• Discuss potential challenges and limitations.
4. Prompt Optimization Strategies:
• Explain the role of prompt optimization in AI, particularly for large language models.
• Explore the connection between prompt optimization and meta-learning.
• Showcase examples or best practices for designing and refining prompts for improved AI performance.
5. Synergy and Future Research Directions:
• Explore potential synergies when integrating open-endedness, meta-learning, prompt optimization, and evolutionary algorithms.
• Speculate on future advances: how might these concepts evolve with breakthroughs in AI architectures or computational power?
• Suggest real-world applications (e.g., robotics, healthcare, creative arts, etc.).
6. Conclusion:
• Summarize key takeaways.
• Emphasize the importance of continued research and experimentation.
7. References (Optional):
• Include references or key academic articles for further reading.
• If referencing specific works, provide basic citations (Title, Author, Year) or links.
```
i did't prompt that by hand of course, but used my meta[prompter-3000](https://poe.com/prompter-3000x335)
---
// 31 mar 2025
#metalearning #training #in-context_learning