Our Year of AI Debates: 2025

Throughout 2025, we hosted bi-weekly AI meetings to provide a forum for TheGP’s engineers, founders, and technical leaders in our community to debate the future of AI. While plenty of time was spent on recent releases and benchmark results, the more interesting moments came when the conversation turned speculative. The fact that some of these speculations now seem obvious reflects how fast things moved this year. Here’s a look at some of the most provocative ideas that emerged:

Winter

January to March Highlights

“The Telnet days of AI interfaces”: That’s how Figma’s Head of AI Products David Kossnick framed the moment in January; back to the command line, not quite knowing what will happen until we hit enter. Chat interfaces lower the floor for beginners, but engineers and designers in the meeting wanted higher-resolution interfaces to raise the ceiling for experts.

The code editor is dead: In February, Charlie Labs founder Riley Tomasek predicted a hasty end to editor-based coding. Soon, he noted, there would be “very little time spent actually editing characters, and a high number of things running in parallel asynchronously.” Releases from Charlie and others have since pulled that future closer, with many engineers already spending less time in traditional editors.

MCP’s uncertain ascent: In March, only a handful of companies like Cloudflare had launched first-party MCP servers. We discussed its takeoff potential and whether it would suffer a similar fate to earlier ecosystem plays from the frontier labs. TheGP’s Alec Flett called it correctly, arguing that being “MCP-first” would soon be equivalent to being “API-first.” MCP has since exploded, with over 10,000 public servers and 97 million SDK downloads.

Spring

April to June Highlights

Claude’s context victory: Southbridge founder Hrishi Olickel joined us to break down his decompilation of Claude Code, arguing that superior context management gives Claude Code an edge. It treats conversation history as mutable, and “the fact that you think you’re having a conversation with the model is almost an illusion.” His takeaway: “Model intelligence has been good enough for a while. It’s tooling that needs to get better.”

Slopsquatting: As vibe coding took off, a new attack vector emerged: AI assistants hallucinate predictable package names, and attackers register those phantoms with malware inside. Agentic tools that auto-install packages compound the risk. As one participant noted: “Once the agent is doing it for you, you’re like sure, go install it.” Engineers should mind the gap between “it works” and “it’s safe.”

Voice agents in healthcare: When Syllable rebuilt their voice platform with LLMs, they eliminated 99% of their codebase, from hundreds of lines of code down to a lightweight core. “The prompt is the guardrail,” their VP Product Ibrahim Cotran said in the meeting. “We’ve done millions of calls now, and we haven’t had a problem.” Industry-wide, the momentum is in solutions like Syllable’s: administrative, fast to deploy, and far from patient risk.

Summer

July to September Highlights

Coding agents eat the market: An agent that can write great software can do a lot more. Participants described using coding agents for tasks beyond coding: product planning, task management, marketing, internal documentation. The question might not be whether specialized agents will emerge for each domain. It’s whether coding agents will simply subsume them all.

Gold medals without the crutch: OpenAI and Google both claimed gold-medal performance for the 2025 International Math Olympiad—this time without translating problems to formal proof languages like Lean. If models can skip formal language scaffolding to solve IMO problems, how many other multi-step processes can be collapsed?

The RL economy’s half-life: The infrastructure fueling reinforcement learning (RL), from Mercor to RL environment startups, faces a durability question. Once an agent masters a skill, the training data’s value declines, along with the humans who generated it. One participant put it bluntly: “The value of the labor pool will depreciate. No question. Will it depreciate too fast so that they can’t get an acquisition beforehand?”

Fall

October to December Highlights

Don’t reinforce, iterate: RL dominated the conversation this year, but it remains difficult to execute. One participant asked, “Can we skip all that infrastructure buildout and PhDs and instead just ask it to make its instructions more aligned with our desires?” This approach doesn’t always work, but participants noted success with tools like Claude Skills, which self-modify instructions based on feedback.

Coding agent failure modes: Engineers in the meeting who embraced async coding agents are now facing a review bottleneck. Alongside review tools, they often rely on a growing intuition of where certain agents struggle. TheGP’s Dan Pupius noted he now spends the majority of review time on the small fraction of AI-generated code he knows is most likely to contain errors.

Revealed intelligence: Most AI assistant products first engage users with polite, professional interactions. Poke took the opposite approach. In onboarding, it might goad a user into granting it data access, then wield that data in price negotiation. Reactions ranged from beguiled to repulsed, but everyone agreed that Poke’s sardonic style did a much better job exposing its smarts than a more buttoned-up approach.

These bi-weekly conversations are always a highlight at TheGP. We want to thank all our guest attendees, including David Kossnick, Riley Tomasek, Hrishi Olickel, Ibrahim Cotran, Tyler Fonda, Mike Adams, Sami Torbey, John Smart, and Michael Magan for joining. If you’re a founder or builder working on AI and would like to join these debates in the new year, please reach out.

And thank you to our team at TheGP for always keeping these timely debates lively and thoughtful: Alec Flett, David Watson, Mikey Wakerly, Stas Baranov, Justin Rosenthal, Marcus Gosling, and Ted Mao.