Pichai Tempers RSI Talk: Behind Argon, Researchers Still Run Experiments with Agents

On the Big Technology Podcast, Alex Kantrowitz asked Sundar Pichai about recursive self-improvement. Pichai did not dodge the question: RSI was a key contributing factor in developing Argon, but it is still far from the endgame of RSI. A more accurate description of the current state is that there is an RSI loop alongside researchers, with agents helping researchers build these models through automation.

Let's clarify RSI first. Recursive self-improvement refers to AI participating in improving the next generation of AI, forming a research loop. It sounds like automatically writing code, automatically running experiments, automatically evaluating, and then automatically revising. Pichai's account breaks that loop down: agents are helping researchers automate part of the work, and researchers are still in the loop. Someone pressed on whether agents account for 5% or 60% of the work on Argon, but the podcast did not give a number. Without a number, the statement can only be a directional judgment.

Argon itself is the backdrop. After Gemini 4 Argon launched, Pichai said it used nearly 25 trillion tokens within 72 hours, roughly three times the same period for Gemini 3, and that service capacity was constrained. Big Technology's report noted that Argon surpassed the models OpenAI and Anthropic had publicly available at the time on coding, math, science, and knowledge-work benchmarks. Pichai also acknowledged that coding still has pockets that need improvement. That detail is more useful than the slogan of 'the strongest model.'

Compute allocation is another thread. Pichai said Google Cloud has at times given up short-term revenue to hand compute over to frontier AI development. This trade-off is not abstract: with the same batch of chips and data centers, selling to cloud customers recognizes revenue, while leaving them for DeepMind to train and run inference on Argon burns costs first. Some commenters asked in replies how Google decides which chips make Cloud money and which go to DeepMind. The podcast did not lay out an allocation formula, but Pichai called frontier AI 'ground zero.' For readers with a development background, this means model capability, product experience, and cloud revenue are all placed on the same resource spreadsheet.

One observation: RSI is not the moment when a switch suddenly flips one day. It is more like the automation ratio moving upward bit by bit. Agents first take over the tedious parts—experiment management, log analysis, benchmark runs, code patches—while researchers hold on to direction, evaluation, and failure diagnosis. Only when agents can reliably propose verifiable new hypotheses and then design experiments to falsify them themselves does it make sense to talk about a closed loop. For now, Pichai's phrasing is a loop alongside researchers, and that 'alongside' is crucial. Don't mistake an agent demo for RSI being complete. And don't assume that because the CEO says 'far from the end state,' nothing has happened. Argon's token consumption and changes in internal development efficiency already show that agents have entered the production process.

Some commenters said this was a rare honest answer from a CEO, without a victory lap. Others cautioned that 'far from the end state' is CEO-speak meaning 'we are far from done, and we know it.' Both readings hold up. Pichai gave no percentage and no timeline, only the current form. For people following AI progress, that is more informative than a line like 'AGI is coming soon.'

The full podcast is on Big Technology Podcast; YouTube link: https://youtu.be/F3ItSTZNtT0 . The transcript is on Big Technology's Substack, with some content behind a paywall. Original tweet: https://x.com/haider1/status/2114219149602050144 .