GPT 5.4, NVIDIA GTC, AI Impact on the Job Market | Ep 12
TL;DR
GPT-5.4 feels like a real upgrade over 5.3 for coding and planning — Adam says it’s “so much better” and less robotic, while Ray now prefers GPT-5.4 over Opus for planning, especially in Cursor with high or extra-high reasoning.
More reasoning isn’t always better — the hosts found GPT-5.4 on extra-high can overthink, hallucinate, and “spin out,” echoing old o3 behavior, so medium reasoning often works better for focused repo edits.
Huge context windows are useful, but compression still wins — despite Opus getting 1 million tokens in Claude Code and Cursor, Ray says the sweet spot is often closer to 80k tokens, because once you blow past 200k, noise and duplicate context start degrading model performance.
NVIDIA’s GTC message was clear: inference is the business now — Ray’s biggest takeaway from the conference was that Nvidia wants to own the stack “from the electron up to the app layer,” with newer hardware positioned around inference efficiency, throughput, and margin expansion for customers.
The real AI job-market story is still murky, but the work is undeniably changing — the hosts push back on simplistic “AI replaced workers” narratives, arguing many layoffs are really overhiring corrections, while technical writing, agency website work, and repetitive white-collar tasks are already being hit hard.
Personal agents and ‘Open Claw strategy’ are becoming a real product question — instead of treating it like a meme, they frame it as a practical challenge: make software usable by agents through CLIs, MCP, skills, and better context/tool routing.
The Breakdown
GPT-5.4 finally feels less robotic
The episode opens with the crew catching up after a weather-related delay, then jumping straight into GPT-5.4. Adam says it’s “so much better than 5.3,” mainly because 5.3’s robotic tone was such a turnoff, while 5.4 feels smarter and more pleasant to work with. Ray says he now prefers 5.4 for planning over Opus in many cases, especially inside Cursor, where it shines on tool-calling and repo-context tasks.
Why Cursor + GPT-5.4 clicks for refactors
Ray describes using GPT-5.4 in Cursor to do a major repo refactor and add CLI capabilities for Open Claw workflows. His pattern is to use extra-high reasoning for planning and medium for more targeted edits, because extra-high tends to do more loops, pull more files, and behave like the old o3 model he loved — fast, capable, and great at research. The catch: it’s expensive enough that he jokes it feels like being a “tier one engineer with infinite credits.”
The trap of overthinking — and over-planning
That extra reasoning comes with a downside: Eric says GPT-5.4 on extra-high can hallucinate more, over-resolve insignificant details, and basically get in its own way. The conversation then shifts to plan-heavy workflows, where Eric cites a post from Dex Horthy arguing that if your spec is as detailed as the code, you’re just reviewing a lossy version of the implementation. Adam pushes back a bit — he likes reviewing a one- or two-page plan for high-level decisions — but everyone agrees that 97-page AI-generated spec docs are pure “spec slop.”
Bigger context windows sound great, but they add noise fast
When they move to Anthropic’s 1 million token context announcement, the mood is skeptical rather than dazzled. Ray says he hasn’t seen much benefit from massive windows in practice and prefers the automatic compaction in Cursor; for him, Opus often performs best around 80k tokens, not 200k+. Eric explains why: long agent loops create duplicate file versions, stale edits, and ambiguity, which degrade model intelligence unless the context is carefully organized up front.
GPT Pro and long-running deep review workflows
Eric also highlights a practical change on OpenAI’s side: ChatGPT Pro can now take much larger prompts on the web — around 160,000 tokens instead of the old 60,000 cap. That enables deep code reviews, optimization passes, and long-running research sessions that can take 45 minutes to an hour. His point is that the “Pro” tier behaves like a real agent now, especially when paired with tools like Repo Prompt for exporting structured context.
On the ground at GTC: Nvidia everywhere, inference over everything
Ray’s report from Nvidia GTC is that the whole event felt enterprise-heavy and Nvidia-centric in a way he didn’t fully appreciate before. He says Nvidia is embedded “from the electron up to the app layer,” spanning power, infrastructure, software, robotics, and AI, with the big strategic push now centered on being the “inference king.” The hosts also talk about Vera Rubin, Groq, throughput-per-watt, KV cache workflows, and the idea that the latest hardware can massively improve margins for companies already serving AI at scale.
DLSS 5 and the weirdness of AI changing the image itself
The gaming detour is one of the more vivid sections. Eric describes DLSS 5 not just as relighting, but as subtly redrawing frames in a more photorealistic style — which is why it starts to feel uncanny. Adam says some clips looked incredible, but in others the characters barely looked like the same people anymore, and frame-by-frame inspection still showed classic artifacts like distorted objects in the corner of the screen.
Open Claw, personal agents, and the identity crisis of AI work
From there the conversation widens into “Open Claw strategy,” which they define as making software usable by agents via CLIs, MCP, and agent-friendly workflows. That leads into a more emotional segment about Mo’s viral “sausage” metaphor: software work shifting from artisanal craft to assembly. Adam says he now cares more about designing the best sausage than handcrafting every line, while Ray talks more personally about craftsmanship, identity, and the fear of losing the thing that made you feel useful and able to provide.
Job displacement is real, but the timeline is still contested
The final stretch turns to Andrew Yang’s warning that AI could displace 20% of white-collar workers. The hosts agree some roles — technical writing, website agency work, repetitive office tasks — are already being squeezed, but they push back on the idea that current layoffs cleanly prove AI replacement; in many cases, they see overhiring and executive positioning dressed up as inevitability. Adam’s “canary” metric is telling: OpenAI still has roughly 630 open roles, so until the labs themselves stop hiring, he’s not ready to declare full labor collapse — though everyone agrees the uncertainty is the scary part.