Q2 2026 funding rounds, an Anthropic essay on recursive self improvement, and a new wave of automation focused labs suggest the next 'strong get stronger' axis is AI that helps design its next generation.
Coding-agent competition used to be the easy headline for AI in 2026. OpenAI and Anthropic were trading blows on the developer tools that write and edit software, and the pricing moves that came with them made for clean weekly copy. By late June, that storyline was already feeling thin. The Q2 frame, captured in the LatePost tech podcast LateTalk 171 with Henry Yin, founding partner of MoE Capital, is sharper. The new axis of "strong get stronger," in the LatePost host's framing, is no longer who sells the best coding agent. It is who can build a machine that meaningfully improves itself.
That pivot is not just a podcast host's read. Anthropic published a long-form essay in late June, "Recursive Self-Improvement", laying out a research agenda for systems that help design and train the next generation of models. RSI, the formal name for the bet, is the work of making AI take on some of the labor of making AI better, so the marginal cost of the next capability gain falls over time. The essay matters less as a research breakthrough than as a marker: a frontier lab is now putting RSI on the public map as the thing to watch.
The funding tape in the same two weeks backs up the marker. UK-based Recursive raised at a reported $4.65 billion valuation with Nvidia and GV participating, a step up from the seed-stage valuations that used to define the RSI cohort. A team of ex-Anthropic researchers closed a roughly $200 million seed at around a $1 billion valuation within weeks of leaving the lab, a number that prices pedigree and timing, not product. Jerry Tworek, a former OpenAI researcher behind the original GPT-4 work, launched Core Automation with the stated goal of building "the most automated AI lab in the world." Each name is a different bet. They cluster on the same axis.
That clustering is the part the wire stories do not connect. Funding announcements read as a parade of unrelated rounds. Read together, they describe a market that has decided RSI is the next category worth funding, and is pricing that decision before most of the teams have a public product. The LatePost hosts note that several additional RSI-focused teams have surfaced in recent weeks, with some still in stealth, the underwater portion of the cohort that public reporting cannot yet name.
Why does this favor the frontier labs? The LatePost framing is direct: RSI depends on training data, compute, and a research loop that can absorb self-generated outputs, the same assets Anthropic and OpenAI already hold in volume. The next tier (Recursive, Core Automation, the ex-Anthropic seed, and the still-stealth teams) has at most two of the three, and usually only one. The "strong get stronger" pattern that defined the 2024-2025 model race does not flatten when the race moves to RSI. The argument from the LatePost discussion is that it sharpens, because the inputs that drive RSI are exactly the assets the leaders already own. Henry Yin, the MoE Capital partner on the episode, treats this as the natural extension of a trend rather than a disruption of it.
Two second-order threads from the same Q2 cycle reinforce the framing. First, robotics: Sam Altman officially announced the OpenAI Robotics team in late May, and Anthropic is reportedly considering its own push, according to the LatePost discussion. Physical AI is the testbed where RSI gains compound fastest, because real-world data is scarce and self-generated robotics data is the obvious substitute. Second, enterprise own-model demand: companies that want to fine-tune or operate their own frontier models are a joint opportunity for U.S. infrastructure providers like Fireworks and Applied Compute, and for Chinese open-source models like Zhipu's GLM, per the same discussion. Both threads are downstream of the same dynamic. If frontier labs can use RSI to widen their capability lead, the gap they sell into enterprise and robotics buyers widens with it.
The honest limits of this argument matter. The Anthropic RSI essay is a research agenda, not a benchmark result. The funding rounds price a thesis; they do not yet prove the thesis converts to product. Independent replication of RSI gains in open models has so far been modest, and the most aggressive claims about model self-improvement have a track record of softening once other labs try to reproduce them. The trade-press framing that places Anthropic as "confronting the RSI clock" is closer to the right register: an industry under time pressure, not one that has solved the problem.
What to watch in the back half of 2026: any public RSI benchmark from a frontier lab with reproducible gains, the first product-shaped disclosures from the stealth RSI cohort, and whether the OpenAI robotics team can translate RSI research into something a customer can deploy. If those land, the Q2 framing holds, and the coding-agent pricing war of Q1 will look, in retrospect, like the warm-up act.