melvinmelih

This paper has been available for a few weeks, and I wrote an article [1] exploring how to apply its inner workings to the design of multi-agent systems. If you can design "reasoning" at the model level, you can also design "reasoning" in larger, more complex systems using the same principles.

https://melvintercan.com/p/lessons-from-reasoning-designing

> Zhiyuan Zeng, Qinyuan Cheng, Zhangyue Yin, Bo Wang, Shimin Li, Yunhua Zhou, Qipeng Guo, Xuanjing Huang, Xipeng Qiu

I'm guessing English isn't their first language.

You can also subscribe to arxiv email notifications directly, but since there’s 20-30 AI papers coming out per day, it can be a bit overwhelming.

Instructions: https://info.arxiv.org/help/subscribe.html

> but they never solve the first problem I have, which is actually constructing the KG itself.

I’ve noticed this too and the ironic thing is that building the KG is the most critical part of making everything work.

I speak Dutch fluently (born and raised) and even I have a hard time to speak Dutch with Dutch people. If you don't fit the profile (blond hair/blue eyes) they automatically assume you're a foreigner.

One of the most life altering essays I’ve ever read is Montaigne’s To Philosophize Is To Die (https://hyperessays.net/essays/to-philosophize-is-to-learn-t...) where he lays out the principle of memento mori (“remember to die”). Fear of death is often very debilitating, and a topic we all like to avoid but we all have to deal with it, sooner or later. The sooner you accept it, the freer (and happier) you’ll feel.

While I agree with the premise of keeping it simple (especially when it comes to using opaque and overcomplicated frameworks like LangChain/LangGraph!) I do believe there’s a lot more to building agentic systems than this article covers.

I recently wrote[1] about the 4 main components of autonomous AI agents (Profile, Memory, Planning & Action) and all of that can still be accomplished with simple LLM calls, but there’s simply a lot more to think about than simple workflow orchestration if you are thinking of building production-ready autonomous agentic systems.

[1] https://melvintercan.com/p/anatomy-of-an-autonomous-ai-agent

Interesting that he left out the concept of safety (being the founder of a company called Safe SuperIntelligence). Would have been curious to his thoughts on that.

No mention of Perplexity yet in the comments but it's obvious to me that they're targeting Perplexity Pro directly with their new Deep Research feature (https://blog.google/products/gemini/google-gemini-deep-resea...). I still wonder why Perplexity is worth $7 billion when the 800-pound gorilla is pounding on their door (albeit slowly).

Yes, I do offer a premium “Becoming the Author” option that allows you to publish the book under your own name, with full copyright, and the book will be taken offline from the website.

Currently using GTP-4o combined with Perplexity API context infusion for real-time knowledge (this also reduces hallucination for the most part). And yes, starting with outline and then writing chapter by chapter while doing real time research. After initial completion there’s another editing round to make everything more coherent.

Think of it as deep diving into a topic that you want to know everything about. For example quantum computing, you can generate a book that goes in depth in yesterday’s quantum chip announcement (https://instabooks.ai/products/df1ae367-e0c1-b749-9898-2744b...). Or if you’re preparing for a job interview and need to brush up on your knowledge in that specific field (https://instabooks.ai/products/041ced4c-efed-92e4-7ae6-9ad2b...)

Thanksgiving last year, after GPT-4 was released, I realized large language models were finally good enough to bring my idea to life: an AI book generator. Over the holiday weekend, I built the prototype for https://instabooks.ai, to allow anyone to instantly generate 200+ page books on any topic. I sell them in pdf, epub and print form and they take about 30 minutes to complete.

Since launch, over 10,000 books have been generated on topics ranging from niche hobbies to advanced research. The system runs almost entirely on its own, requiring only occasional updates and customer support. The best part of it: as new LLM models come out, the books get better written as well, so each year it improves without much effort on my end.

Depends on which universe you're talking about.

> It performed a computation in under five minutes that would take one of today’s fastest supercomputers 1025 or 10 septillion years. If you want to write it out, it’s 10,000,000,000,000,000,000,000,000 years.

If it's not, what would be your explanation for this significant improvement then?

This is great but will be DOA if OpenAI (80% market share) decides to support something else. The industry trend is that everything seems to converge to OpenAI API standard (see also the recent Gemini SDK support for OpenAI API).

Why .site?

urlshortenersaresoyesterdaytrythisamazingsuperlongurlexpander.com is still available

In case you want to try it out without hassling with the API, I've set up a free tool for it so you can try it out on WhatsApp: https://instatools.ai/products/fluxprovisions

This is a valid concern and while there are no guarantees that the AI won't hallucinate (hence the disclaimer in the book, especially for medical topics), I try to minimize it by pairing the writing with real-time research from Perplexity, so at least it is (or should be) based on verifiable information.

As a Turkish speaker, I really like the name!

“You could parachute Sam into an island full of cannibals and come back in 5 years and he'd be the king.” - Paul Graham

I agree, but that’s what you get when your mission is AI Safety so it’s going to be a dull experience.

I suspect very little people joined OpenAI for their noble non-profit mission after they introduced their for-profit subsidiary. OpenAI compensation was and still is top notch. Compare it to Signal, which is a true non-profit (and salaries are a lot lower).

LA is big and has a lot of traffic for sure, but most adjacent neighborhoods can be reached within a 20 min drive. It’s only when you want to go from one side to another that it will take you more than an hour.

> But it became popular almost instantly after launch, and grew organically from there, reaching millions of daily users.

The law of big numbers dictate that if there’s even a tiny chance of a catastrophic event it has close to 100% probability of happening if n is just large enough (in the case of millions of daily users, probably multiple catastrophic events per day). This kind of asymmetrical risk is very hard to defend against no matter what you do.

When you both follow this advice, it leads to both talking for 30%, leading to awkward silences for about 40%.

Just kidding, but it’s generally true: two introverts have not much to talk about. I certainly always enjoy talking to extraverts because they fill up the silent space.