It's already been a few times when I was forced out of the crypto market because I didn't end up having money to keep on building anymore. At all times when this happened, the outlook felt very uncertain and completely unmanageable. At the end of the day, however, it always turned out to be a mistake to stop working in crypto and find a way to work elsewhere instead. Had I continued to manage working in crypto on crypto-native things, I could have caught the early ICO boom in 2015. I could have also caught the Ethereum ICO. I could have been one of the first people to work with Solidity. And in later cycles, I could have co-invented the Layer 2.

Something to keep in mind here. I don't think we can continue to rely on the Kiwi community to fund ourselves or even to justify working on the product. The market has been so bad that despite us having a decent product and a decent team, the app has not taken off.

This is fine because market conditions can change dramatically over short periods of time. I have seen this countless times, especially in crypto.

We basically invented the NFT in 2015, failed to receive funding, and then right after we failed in 2016, 2017, and 2018, basically an entire ecosystem of projects was funded and hence spawned the NFT craze that climaxed in 2020.

I don't think it is truly beneficial to us to hold accountable users, funders, or even the ecosystem. The fact of the matter is that the hype is in AI, and rightfully so, because with every new release I am once more impressed about what can be achieved with this technology. Why would you embark on an ideological pilgrimage into the crypto space when there are so many other exciting things you could be doing instead?

The fiat sovereign individual will cry wolf to the politician to fix the situation, to ask for bailouts. Why didn't politicians fix this or that? Yet there are so many crises which cannot even be fixed with politics. Self-sovereign individualism also means that we have to be self-sovereign in regards to the beliefs that we choose when we set out to do something. I cannot expect my state, my community, or the Ethereum ecosystem to keep on supporting me. There is no foundational framework of morals that would privilege me to survive in this chaos over anybody else.

I know that what I'm doing is beneficial to everyone. And I know that my freedom to do whatever I want keeps me mobile in the face of increasingly hostile technological powers which know everything about us. A decentralized social network is potentially conceived on the idea of trading off individuals' privacy for communal transparency. Yet it turns out that the social contract of what data can be used for is nowadays just kicked by the feet of powerful companies in California.

The beauty of self-sovereignty in entrepreneurism is that I don't owe anyone any favor. I can choose to change my mind every day without talking to investors. In fact, even without talking to my co-founder, since we, too, are individuals, too. And so this makes us response-able because it allows us to pivot and to fill voids created by technical revelations.

The reality is that tool call will make LLMs so powerful in a not distant future that every information on the Internet will be accessible through simple prompts. Without going into much detail, you can already prompt for how the salt and pepper meals in your favorite restaurant look like. You can ask O3 what the most common complaints are about certain waiters. So you will be able to learn everything about everyone, anywhere.

For a long time I was arguing in this community and elsewhere for differential privacy, conditional privacy, and against privacy maximalism. However it is starting to look as if encrypting vital metadata and content is going to become an imperative.

You may say there is no market for us, but that we are a good team and that we have a decent product. But what if we actually don't have a very well-fitting product for a new type of market? And what if we actually have the freedom to build a social network with privacy implemented at the core protocol?

I always find it funny when people get almost blinded by AI.

For my understanding, it's just that blockchain is kinda boring. Like animated GIFs compared to AWS. Or CSS compared to kernel development. A bit exaggerated, but you get the idea...

Blockchain is by an order of magnitude more impactful than AI. Bold statement, I know, however let me explain:

The world is apparently just going nuts over the impact of the digital Yuan compared to the SWIFT network. They hype up how great it is that smart contracts are releasing funds when goods arrive in a port instantly, without waiting 72h on SWIFT. And not spend about 6-10% in transfer fees. That saves some $30T in trade fees. Well, a little fairy told me that technology exists somewhere for about 10 years now... (disclaimer: don't know the guy, got the post forwarded, don't know how credible the content really is, but it made its rounds, that's enough for me to understand the hype is real, so here it is the source: https://x.com/kingkong9888/status/1914081678143767034?s=46)

PwC estimates AI will generate some $16T in contribution to the global economy by 2030. (https://www.pwc.com/gx/en/issues/artificial-intelligence/publications/artificial-intelligence-study.html). I think the other large consultancy firms estimate that in the same ballpark, some $20T some less. This is now a hyped up number, I firmly believe the actual number is smaller, because they know incorporate all kinds of "if this then that" wonderful scenarios which will never happen.

I also already see some of the blockchain-turned-AI-fluencers slowly coming back to blockchain.

So, if I put this into perspective: hyped up AI will maybe generate $15-20T until 2030 (5 years). Blockchain still can build efficiency today and save $30T. The big consultancies will push this message at some point and it will have ripple effects (no pun intended). Blackrock wants to bring equities trades on the blockchain and there's really just one contender (its not the EU CBDC and definitely not the digital yuan either).

I think the future is bright for blockchain developers, especially the ones building on the EVM.

Also, interestingly: LLMs are great and summarizing. They make large text corpus small. They are a real shitshow expanding text. They are not thinking, they merely run token weights, they can't pull information out of thin air. They are great building JavaScript pages, or python scripts, because they had so much samples to learn from. They are a shitshow doing proper solidity/smart contract architecture. I know. I tried. It's hilarious. Ask gemini-2.5-pro about something even slightly more complex than an erc20 token if you need a good laugh. And since that is such a niche which is in emerging development, they will lack behind for a long time...

My 2 cents...

Can LLMs fundamentally be trained to be better at writing Solidity than humans? Elon Musk says cars can be made safer than humans. But you just need one humans to break all Solidity contract agreements of many people, so it‘s not comparable. So we cannot ever train a Solidity LLM and make it operate autonomously?

I mean ... "it depends"?

LLMs are great at automating tasks away.

They are great when standardised procedures are done over and over again and it can rinse and repeat.

Like CRUD functions. Like the bazillionth middleware for express. Like a relatively standard erc20 token. Like the regular SQL query.

Try doing something where it would need to really "think" and come up with something, it can't.

So, I guess, it depends if the "Solidity LLM" can burp up code that's been written a billion times already somewhere. Then it creates efficiency. And I use it every day for that. I can write bad code faster now (/s).

I don't think that's the fundamental comparison for solidity versus other tasks which are in the sample set of things the LLM has learned in training. Because even if the LLM was to regurgitate the most standard ERC-20 contract, the failure threat is that millions or billions of dollars are going to be stolen from potentially hundreds of thousands of people. Whereas when I drive with my Tesla and something goes wrong, the worst case is that me and my family, we lose our lives. But you can also always blame it on the driver who didn't interrupt the autopilot. And at the same time, one failure in an autopilot wouldn't result in damage to all autopilots in operation.

So somehow I think there is a difference in using the perfect solidity LLM to write smart contract code versus using an autopilot to drive a person's car. Because technically you could make a solidity LLM extremely good at writing safe solidity smart contracts as somehow there would be a way to do reinforcement learning. Imagine if you have one LLM trying to hack another LLM's contracts and you scale this up to some crazy computational degree where the LLM would become extremely good at writing safe solidity smart contracts, potentially better than any single human could do similar to how the Tesla autopilot could become a better driver than any single human.

Because at the end of the day, even though your LLM may write better solidity code than any single human, somehow as the human you still carry a responsibility for delivering the solidity to the user base who is going to store money in it. And then there are dynamics around creating insurances and spreading liability among many different humans. Which somehow are different in quality than having the contract being reviewed by several really smart LLMs, all individually smarter than humans. Does this make sense?

Hm, I think I get what you mean. I am using it to regurgitate the boilerplate code, but then go through everything manually again to make sure it works. Bottom line is, while I need to bend it a lot to make it do what I want, it still is more efficient than me writing the all the code in the first place. And for the last mile I code it myself.

But one situation was particularly odd.

I recently had an argument with Gemini.

I am writing a set of Contracts that handle overcollateralization of liabilities denominated in different other ERC20 tokens. What they exactly do is not really important, other than the liabilities and the collateral are both fungibles resulting in an NFT that represents that overcollateral as a tradeable position. Anyways, here is where it gets ridiculous.

I know its underlying asset is fungible. I am 100% on that. I had to argue my way with Gemini because it simply didn't believe me.

After going through all the arguments why it thinks everything is non-fungible and this is still fungible underneath, the last argument from Gemini was: "But this feels wrong because User B's position hasn't changed."

It was one of these *scratch my head* moments where I also don't know what to say anymore. Like "Yeah you are right, but I don't believe you" situation.

Anyways, I would not advise to use any AI for coding an autopilot without human supervision. Not on the road. Not in aviation. Nowhere.

tbh it‘s been a while I really had an unaligned LLM actually. I feel like ever since I‘ve become extremely precise with my aider conventions file they do what I want