Vibe coding is brilliant for developers who are able to accurately define requirements and provide expert correction to LLM-generated outputs.
But any platform that democratizes custom software development for everyone else must meet a much higher bar.
This is why since our inception we've built software for non-developers via a spec-first interface. It's also why our generator builds software block-by-block by consulting a world model of trusted code recipes.
This was a super fun chat with Antonio and Upal of
@bem_hq.
We discussed why building a product that creates and maintains working software apps from natural language specifications is hard, and how we're approaching the UX and the tech that enables it Durable.
Happy New Year!
We need good software products, not more code. Code is just the means to and end, not the end result. Codegen as a generator of tokens will stagnate as a novelty, unable to create products that are scalable, reliable, and solve human problems.
In our first 2025 episode of Hard Software we chatted with @n_keivan, Cofounder & CEO of @durable_ai. They're approaching "codegen" from a perspective I think will win in the long run: 1) Focusing on the entire development lifecycle, from requirements to deployment, and 2) Not just spitting out tokens hoping they'll compile, but focusing on neurosymbolic AI, individual units of compute that work in tandem to execute requirements.
You can catch the entire episode on our blog: https://t.co/OTLR9mjJN7
LLM hallucinations are particularly bad when generating code, and even more so if the code has to work without modification, as in the case of text->product.
So fixing them is necessary if we're to advanced beyond small-scale code autocomplete.
We've published a new post on our blog describing some of the reasons why we're taking a different approach to building the AI and platform that powers text->product.
To build the platform that powers text->product, we've taken a different approach to AI @durable_ai . It relies on planning over learned world models instead of next token prediction. I've outlined our reasons for taking this more challenging approach in a new post.
We're building text-to-product at Durable, and have a new post from our private blog on the vision. TLDR: custom software will become much more accessible as we progress towards text-to-product, but for that we'll need a combination of LLMs and explicit planning/verification.