Colin Gillingham

The through-line

Colin Gillingham's recurring argument is that AI has collapsed the distance between having an idea and shipping it, and that the people who benefit most are the ones who were never able to build before. He describes himself as "always kind of was slow at it uh always the idea guy but never the guy that could do it always needed to find somebody to help me bring my ideas to life," and says that with AI tools he is "actually like a competent developer now because it's more like putting Lego blocks together" [1]. The proof he keeps returning to is Podstash itself: idea to first paying customer in ten days, marketing site included, with "99 of the Chrome extension" written by ChatGPT [1]. He extends the same logic to everyone else, including people who feel threatened by the technology, and he is consistent that the correct response to disruption is to absorb the tool rather than resist it [1]. He has been asked repeatedly to connect this to his own trajectory through big tech, including Tesla and Mapbox [2][3][5].

On AI as a multiplier for the "idea guy"

The central claim is that the bottleneck in software was never ideas, it was execution, and that the bottleneck has moved. Gillingham never trained formally as a programmer; he did a full stack bootcamp in Europe after Tesla and says he "never got really good at it but got I was okay at it" [1]. What changed is that AI turns coding into assembly: "there's so many idea guys out there and with a little bit of patience with AI and figuring out how to work it makes a lot of those things possible" [1]. He frames the ten-day build as "a kind of a testament to what can be achieved if you know how to use AI to help you kind of 10x your productivity," while explicitly refusing to romanticise it: "I was working like 12 13 hours a day not not to give anybody like the wrong idea like I put a lot of hours into it" [1]. His entry point was pragmatic rather than ideological. He started using the OpenAI playground before ChatGPT existed, out of necessity, because a marketing team was being downsized and there were no marketing managers left to write content [1].

On how Podstash actually works

Podstash is a Chrome extension that turns any blog post, article, Wikipedia page or YouTube video into a five minute podcast episode delivered through a custom RSS feed the app generates, which the user adds to Apple podcast, Google podcast or any other player [1]. The pipeline is a chain of model calls rather than a single prompt: long documents are split into roughly 100 to 200 word chunks so they fit the context window, each chunk is summarised, and the summaries are then fed back with an instruction to write a compelling five minute podcast script [1]. The script goes to 11 Labs for voice synthesis, which he rates as "the best I've ever heard of like a text to speech technology," and the output convincingly imitates the form, opening with "hello listeners welcome today episode is about" [1]. An optional translation call plus 11 Labs' multilingual pronunciation covers French, Italian, Spanish, polish, Hindi and Portuguese, which he sees serving two audiences: non-English speakers locked out of English-only content, and language learners who have exhausted textbooks and want to hear crypto, financial or medical terminology spoken in their target language [1]. He built it partly for himself, being "slightly dyslexic" and disliking reading, and for people with a graveyard of open Chrome tabs they intend to read and never do [1]. Pricing is a two-week free trial then seven dollars a month for up to five episodes per 24 hours, and he is candid that he is "still trying to kind of figure out like what my costs are" given the stack of API calls behind each episode [1].

On following the signal from users

He treats unintended usage as data rather than noise. About half of Podstash's early signups turned out to be podcast hosts, bloggers and industry curators who wanted to build a show out of what they found interesting online, which was not the use case he designed for [1]. His conclusion is a general one about products: "whenever you build a product you kind of like take a step back and watch how your users are using it and then sometimes you're surprised how they're using it because that's not how you intended it uh but you you know you you can't ignore that you you have to kind of like follow the signal" [1]. Those curators are also his most vocal customers, so the roadmap bends towards them, including a possible ChatGPT plugin where a creator researches and drafts inside ChatGPT, then sends the script to Podstash and ends the conversation with a new episode published to their audience [1]. He was shipping a major update roughly every week at around 150 customers, three weeks after launch, on the basis of that feedback [1].

On the coming flood of low-quality AI content

He accepts the criticism without accepting the conclusion. When he pitched the idea to Jason calcanus of the all in podcast, the immediate reaction was that he would "fill up the internet with like trash AI generated podcast episodes," and Gillingham grants the underlying fear while denying it describes his product [1]. He points at SEO as the existing version of the problem, where tools take a keyword list and produce 500 word posts, and he admits doing it himself: an entire blog of about 20 posts on podstash.ai in three hours, built expressly to capture Google search traffic [1]. His expectation is that this is a phase rather than an end state. There will be a burst of noise, comparable to mid-journey images with "15 fingers," followed by adaptation: "there will be like uh a filtering system that will kind of filter out the low quality stuff," with Google deprioritising and reranking accordingly [1].

On what LLMs can and cannot invent

He draws a sharp line between fluency and novelty. Today's models are "generalized large language models they're not tuned for anything they're just generalized they're like okay at everything fine," and the next step is fine-tuned variants, for instance a base model trained on every movie script ever written, which he expects to be genuinely good at writing scripts [1]. Even so, he holds that language models cannot escape their training distribution: "with the llms their creativity is kind of confined to the realm of all the ideas that have already happened," and "the great thing about being human is like you can you can think of something new that all the other humans haven't thought of yet" [1]. His one counterexample is deliberately drawn from outside that lineage. In the alphago documentary, the machine played a move against the world master in Korea that no human had played or considered good, and won because of it, which he calls "actually a display of machine creativity" [1]. He is careful to note this sits in a parallel branch of machine learning rather than the LLM universe [1]. Whether AI can write good movies and TV, or act in them, he places as "not quite but it will get there" [1].

On striking against the technology

He is unsympathetic to refusal as a strategy while conceding the substance of the grievance. His analogy is a horse and buggy company striking against the arrival of the car: "you're doing it at the absolute worst time to to to stop working" [1]. At the same time he thinks the economic complaints are legitimate, that the business model and revenue sharing are wrong and that royalties need reworking now that everything has moved to streaming, and says "I think they're right that's probably that's probably true" [1]. The prescription follows from his broader position on productivity: rather than banning the tool, "they need to use AI to become 10 times more productive than they are now" [1]. He has discussed the broader question of whether AI can be governed or contained in a conversation billed around exactly that [3].

Takeaways

• The bottleneck in building software has moved from execution to ideas; with AI, coding becomes "more like putting Lego blocks together," which puts shipping within reach of self-taught tinkerers [1].
• Ten days from idea to first customer, with ChatGPT writing about 99 percent of the Chrome extension, but on the back of 12 to 13 hour days [1].
• Long-form content is handled by chunking into 100 to 200 word pieces to fit the context window, summarising each, then generating a five minute script, with 11 Labs doing voice synthesis in up to seven languages [1].
• Half of early Podstash users were creators building curated shows rather than the busy readers he designed for, and his rule is that you cannot ignore that and must follow the signal [1].
• Expect a wave of low-quality AI content, including SEO spam he has produced himself, followed by ranking and filtering systems that push it down [1].
• LLM creativity is bounded by ideas that already exist; the AlphaGo move that no human had played is the counterexample, and it comes from a different branch of AI [1].
• Fine-tuned models, for example trained on every movie script written, will outperform general-purpose ones on domain tasks [1].
• On the strike: the revenue-sharing and streaming royalty complaints are fair, but refusing the tool is the horse and buggy response; the answer is to use AI to get ten times more productive [1].