Aug 6, 2025 · 1h 14m · latent-space
The AI Agenda: GPT5 leaks and the business of AI News — Steph Palazzolo, The Information
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Stephanie Palazzolo, AI reporter at The Information, joins the Latent Space podcast to discuss the investigative craft of tech journalism, the unit economics of AI inference startups, and the rapid evolution of frontier models and autonomous coding agents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 40.5% of the talking time here. How this is scored →
The hosts as informed peer 5.9 Guest teaching 4.1 Guest disagreement 1.1 The hosts pushing back 1.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:12:56 Stephanie clarifies why RL agent sandboxes are hard to scale
Stephanie counters Swyx's assumption that RL environments should work out of the box by explaining the massive compute cost and complexity of simulating full enterprise workflows like Salesforce.
Hardest push from the hosts ▶ 1:12:48 Swyx presses on whether RL environments are workingSwyx challenges Stephanie's assessment that RL environments are still early, directly probing whether the technology is failing to deliver on immediate expectations.
Biggest teaching moment ▶ 55:05 Stephanie details the mechanics of source anonymityStephanie provides a comprehensive insider look at the precise phrasing and sourcing ethics journalists employ to protect startup employees and whistleblowers from getting identified and fired.
The host holds their own ▶ 1:11:00 Swyx explains the Transformer lock-in via the Esperanto analogySwyx demonstrates deep domain fluency by explaining why non-transformer architectures fail to gain traction despite theoretical merits, drawing parallels to the lottery ticket hypothesis and Esperanto.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Inside The Information Newsroom and Daily Beat Workflow | 3 | 3 | 0 | 0 | Swyx connects his finance background to industry reporting, while Stephanie demystifies the newsroom beat by revealing that 80% of her day is spent networking and sourcing rather than writing. | |
| The Economics and Market Dynamics of AI Inference Startups | 7 | 4 | 1 | 2 | Alessio and Swyx dive into the economics of GPU inference providers and open-source video workloads, while Stephanie shares insider data showing that 75% of DeepSeek R1 usage on certain platforms was focused on distillation. | |
| Meta's AI Strategy, Personal Superintelligence, and Talent Wars | 6 | 4 | 2 | 2 | Swyx questions how journalists verify rumored nine-figure compensation packages amid CEO narrative games, prompting Stephanie to detail how multi-source vetting and equity vesting structures are verified. | |
| Frontier Lab Leaks: GPT-5, Anthropic, and Unit Economics | 7 | 4 | 1 | 2 | The hosts cite technical research such as the Chroma Context Rot paper to assess Claude Code's edge and question frontier lab unit economics, while Stephanie discusses GPT-5 testing impressions and talent poaching. | |
| Coding Agents Ecosystem and the Windsurf Acqui-Hire Drama | 6 | 3 | 1 | 1 | Swyx introduces the concept of execu-hires and discusses the breakdown of the traditional startup employee risk-reward pact, while Stephanie explains growing VC fatigue with non-standard acquisitions. | |
| The Business of Tech Journalism, Scoops, and Sourcing Ethics | 5 | 6 | 1 | 1 | Stephanie educates the hosts on journalistic tradecraft, cold calling techniques, subscription business mechanics, and the deliberate wording used in attribution to protect insider sources from retaliation. | |
| Pitching Journalists, RL Sandboxes for Agents, and Future Outlook | 7 | 5 | 2 | 3 | Swyx uses the lottery ticket and Esperanto analogies to argue why transformers remain the dominant architecture and presses Stephanie when she describes RL agent sandboxes as early, prompting her to explain simulation and compute bottlenecks. |
Statements from this episode (18)
Fanelli: Image and video generation is much more open-source driven than language
“FAL especially has done amazing because media is still very open source driven when it comes to image and video generation. In a way, the language is not as much.”
Fanelli: Small venture funds cannot back AI inference startups with $5M rounds
“We haven't done any, but also like we have a smaller fund. And I think all of these inference things are like big money. Investments. Like you cannot do a five million dollar round in like an inference company, you know, you kind of need a good chunk of capita…”
Swyx: AI inference providers are a proxy for the open-versus-closed model gap
“The other thing I always think about from a finance point of view that, Maybe people don't even think, don't think about that closely, but obviously push back if you disagree, is that it's basically a proxy for the gap between open models and closed models. Be…”
75% of DeepSeek-R1 usage at a major inference provider is for distillation
“I know even whenever, like, the DeepSeek R-One model first came out, one of the major, like, startups in the space was, like, yeah, 75% of the usage that we see for R-One is people trying to distill it into, like, Whatever specific use cast use case they want”
Palazzolo: Meta should focus on app integration over frontier models
“Maybe what's actually good for their P&L and their finances is to not focus so much on building these, like, insane, huge, state-of-the-art models, which they've obviously struggled with more recently, but it's to take things that are, you know, 70, 80, like, …”
Early GPT-5 testers report noticeable gains across coding, science, and writing
“And the story that we wrote, we kind of talked about how at least the people that we've talked to who tested it so far have been pretty impressed. They seem to think that it's been, you know, there's been improvements in a number of domains and both like scien…”
Chroma research finds Claude models lead in long-context utilization
“And you know, one thing Chroma released this context rod paper recently about context utilization and the cloud models are actually the best at using kind of like longer context.”
The Information: OpenAI hit $12B ARR as burn rose to $8B
“We had a story yesterday about open AI and how, like, I think they've reached about twelve billion ARR and yeah, but their burn went from like They projected, like, one billion to, like, eight billion or something.”
Palazzolo: Claude Code leads stayed at Cursor only two weeks
“We know that they went there, they were there for, I think, about two weeks, and they came back.”
Swyx: Windsurf will stick around as an active product post-acquisition
“I think Windsurf as a product is going to stick around, and people who really like Windsurf, I mean, I was a Windsurf user for a long time we are going to keep using it because it fills a need, and like, obviously, Cognition bought it for a reason.”
Palazzolo: VCs are growing frustrated with unconventional AI acqui-hire deals
“Talking to investors, I think they're starting to, I think they're kind of over these types of deals. Like, they're like, hey, I mean, like, this is, like, fine, but, like, we don't invest in companies so they can get, like, a weird acquihire situation a coupl…”
Palazzolo: Unbalanced AI acqui-hires are eroding employee-founder trust
“There's a sense that whenever you join a startup, like, you're gonna be compensated for the risk that you're taking on of joining, like, something that's unproven, and then now you're having these situations where, like, yeah, some people do get rewarded for t…”
Swyx: Open-source coding tool Cline is announcing a Series A
“And they're announcing their A today.”
Palazzolo: News aggregators and summarizers are being displaced by AI and influencers
“If you're a company that's more known for, like, summarizing, like, the daily news, or even, like, having some sort of opinion or analysis, like, now that, you know, people on, like, X who are doing that, or you can just use, like, ChatDBT to send you summary …”
Palazzolo: Founders should pitch industry trends rather than individual startups
“Don't think about just pitching your startup to me, but, like, Pitch me on, like, why your entire space is important, and, like, I will write something, like, it's more likely that I'll write something about a trend or, like, a new industry.”
Swyx: PR firms are pointless for early-stage startup pitching
“For the early stage startup founders listening, I think most of your PR teams are kind of pointless unless you really want to get on, like, CNBC. Most, most of us are, like, easily reachable, you know, like we screen tons of requests and it's fine. Like it's l…”
Swyx: Most AI researchers are scaling Transformers, not alternative architectures
“I think most people that I talk to are not seriously pursuing alternative architectures. There are some notable exceptions, primarily together with the Mombard architecture recursal with RWKV. There's like the XLSTM that was created by a separate Hogwriter. An…”
Palazzolo: Frontier AI labs are partnering with RL sandbox startups
“At least the one, the startups that I talked to, a decent chunk of them are already starting to work with, like, the AI, the big AI labs, and, like, these are quite early startups, so it seems like they are very, the AI labs are very excited about this approac…”