AI Digest, Oct 25: Creative Job Losses, Solo Devs Rebuilding Postgres, Small Models

0:02PM Lead ProThis is PM Lead Pro. Today's digest covers creative job losses in the US, labs stepping back from video generation, solo developers building ambitious software, and new frontier and local models.

0:18So if you look at the tech industry right now, it actually looks less like a traditional factory and a lot more like an automated assembly line where, you know, the line workers just got handed the keys to the robots.

0:29Yeah, it's a complete paradigm shift.

0:31Right. So we're diving into a pretty dense stack of sources today.

0:34We've got Joseph Politano's latest economic research, Dan Liu's developer notes, and this week's new model specs.

0:42And the core mission here is really to unpack how frontier AI is just completely rewiring the global economy.

0:48I mean, it's actively replacing legacy jobs.

0:50But at the exact same time, it's giving engineers and SME owners the power to build incredibly ambitious software.

0:57Totally.

0:58And looking at the sheer scale of the job displacement in Pulatana's data, I mean, the shift is undeniable.

1:04The U.S. wiped out like 200,000 creative jobs in just four years.

1:08Right.

1:09And Hollywood alone lost 100,000 of those.

1:11That is a third of their entire sector, which is wild.

1:14Wow.

1:15A third.

1:15It really feels like Hollywood is having its Napster moment.

1:20Like back in the early 2000s when unbundling and free digital distribution just totally gutted the traditional music model.

1:27Are we just seeing the video equivalent of that?

1:30Well, the mechanism is definitely similar.

1:32You know, digital abundance destroying artificial scarcity.

1:35But video media is missing the life raft that music had.

1:39Right, because when Napster hit, musicians could just pivot to live concerts and merch.

1:43Exactly.

1:44Exactly There is no stadium tour for a television writer And the tech sector itself hasn recovered from its 2022 peak of 300 jobs either largely because coding agents are well they flattening the need for huge engineering teams

1:58Yeah, but the real driver here isn't just that AI can make video now, right?

2:03It's that the major frontier labs like OpenAI and Anthropic, they are actually stepping back from our generation entirely.

2:10I mean, OpenAI even shelved its video model Sora.

2:14I know, it's crazy.

2:15Wait, they actually walked away from a product that was blowing people's minds just like a year ago.

2:21Yeah, they looked at the unit economics.

2:23Deploying vast amounts of compute on reasoning and coding just yields an exponentially higher return on investment than generating synthetic movies.

2:32Oh, that makes sense.

2:33So they are actively leaving image and video generation to smaller open weight models because solving complex math or architectural problems for enterprises is just vastly more profitable.

2:43PM Lead ProFor owners and team leads: the hosts say labs are putting compute into reasoning and coding because it pays better than video. Expect the biggest gains in coding tools, not in generated media.

2:56Okay, so if all that raw compute is being redirected away from Hollywood and into coding, what does that actually look like on the ground?

3:05Well, looking at Dan Liu's recent notes, the output from solo developers is frankly staggering.

3:11He highlights Michael Malice building PGRust, which is a total rewrite of the Postgres database into Rust.

3:17Right, and he's not just translating it. He's actually finding core memory and numerical bugs in the original battle-tested Postgres architecture.

3:25Exactly. Rewriting a foundational decades-old database engine usually takes a massive, highly coordinated team years to accomplish.

3:33Yeah and doing it as a solo developer in a fraction of the time I mean that really shows where the new baseline for software actually sits Let me push back on the quality though Sure Because if you look at GitHub right now aren like 99 of these AI enterprise

3:47rewrites just low-quality vibe slop?

3:50You know, code that looks functional on the surface but completely falls apart in production?

3:54Well, if you are judging by the standard output from older models, yeah, absolutely.

3:59But that's looking in the rearview mirror.

4:00What matters right now is the breakthrough in the absolute latest frontier models.

4:05Right.

4:06Take the test runs with GPT-6 Astra.

4:08It was asked to implement the ZSTD compression algorithm.

4:12It passed 99% of registered tests zero shot.

4:16Meaning it got the complex implementation right on the very first try.

4:19Yep, without needing a single example or correction to learn from.

4:22And get this, it cost about $1 in compute.

4:25The barrier to building bespoke, highly complex tooling is effectively gone.

4:30Which perfectly tees up the model radar for this week,

4:32because that zero-shot capability isn't happening by magic.

4:37It's the result of two very different approaches hitting the market at the exact same time.

4:41Yeah, on the heavy enterprise side, we have GPT-6 Astra and 5.6 Sol driving those massive coding leafs we just talked about.

4:48And Anthropic also just rolled out Clawed 5.5, so that's Opus, Sonnet, and Haiku, alongside Fable 5.1.

4:56and so on. They updated them with optimized system props to give developers much tighter, predictable control over multi-step reasoning.

5:03But the other side of that radar is just as critical, and that's the tiny local models.

5:08Sebastian Roschka latest research highlights Qwen 3 0 This is a model small enough to run entirely locally on a standard laptop CPU or GPU But wait how is a model that small doing anything mathematically useful for a developer

5:24PM Lead ProFor a team lead: the hosts' advice is not to use AI only to write boilerplate ten percent faster, but to build custom internal tools you could not afford the engineering hours for before.

5:36It's all through a training mechanism called RLVR, Reinforcement Learning with Verifiable Rewards.

5:42Okay, so instead of relying on humans to rate if code looks good, developers are hooking these small models directly up to compilers, right?

5:51Exactly. The compiler instantly tests if the code actually runs. That gives the model a mathematically perfect, verifiable reward signal to learn from.

6:00Wow, that's huge.

6:02Then they distill, or basically compress, the proven reasoning patterns from massive models like DeepSeek R1 into these tiny local models.

6:11So you end up with frontier-level logic running locally on a MacBook?

6:16Precisely.

6:16The sheer amount of power moving directly to the individual is just wild.

6:20So, what is the practical takeaway for you listening right now?

6:25Don't just use AI to write your standard boilerplate 10% faster.

6:30The game has changed.

6:32Definitely.

6:33Use it to build the custom, high-leverage internal tools that you previously just couldn't afford the engineering hours to create.

6:38Yeah, and that leaves us with a provocative question to explore on your own.

6:42If a solo engineer can build a bespoke compiler or a secure database engine over a weekend for a few dollars in compute,

6:51what happens to the massive SaaS industry when the cost of generating custom, enterprise-grade software for your specific business drops to absolute zero?

7:00Something to think about the next time you pay that monthly software subscription.

7:05PM Lead ProFind more at pmlead.pro: the transcript and sources are on the episode page. If custom software gets this cheap, which subscription would you rebuild first?