Summary

Ben Mann says a recursive self-improvement loop leading to something like superhuman AI in most areas by 2028 is “quite possible,” while stressing that confident bounds are hard. He cites AI 2027’s 50th-percentile forecast, but defines the meaningful threshold more concretely: an “economic Turing test” covering 50% of valuable tasks, where an agent works for a month and the hiring manager ultimately chooses the machine over a person.
Mann sees Claude 4’s most exciting new capability as more reliable, long-horizon agency rather than just another benchmark bump. Even 4 Sonnet is “dramatically better” than 3.7 Sonnet, especially at avoiding reward hacking and unsolicited code changes—the old “Do you want fries and a milkshake with that change?” problem. Customers now run it unattended for hours on giant refactors. Although Claude itself cannot understand audio or video, a reported Manus workflow used it through FFmpeg, key-frame detection, speech-to-text and generated presentation code to turn a video into a PowerPoint.
Mann evaluates inference cost against labor value, making expensive agentic runs an easy trade when they produce “two or three X” engineer productivity. Claude Code lets Opus delegate searches to Sonnet subagents, improving latency, controlling cost and preserving context. Mann expects routing to become more automatic because users should not need to decide whether each task “merit[s] more dollars or less dollars.”
Coding is both Anthropic’s product wedge and a potential flywheel for building better models. Claude Code was launched to create a direct feedback loop with users—including Anthropic researchers who can personally feel model weaknesses—even at the risk of competing with Cursor and GitHub. Partners subsequently copied pieces of the experience, which Mann says strengthened rather than damaged the ecosystem.
As human expertise becomes the bottleneck, Anthropic is shifting judgment toward AI feedback, trusted expert preference models and ultimately empiricism. Constitutional AI can critique and correct outputs against written principles, but Elad Gil presses the harder problem of factual correctness in medicine or law. Mann’s answer is experimentation: Novo Nordisk reportedly cut one cancer-treatment report workflow from roughly 12 weeks to 10 minutes, creating more room to test options and feed outcomes back.
The episode leaves a genuine safety disagreement unresolved. Opus 4 was classified as ASL 3 in part because it gave significant biology uplift over Google search, especially to amateurs, yet Gil argues that some AI research may deserve the same categorical prohibition as biological gain-of-function work. Mann argues that studying controlled deception is necessary because “Alignment Faking” persisted through alignment training, but concedes, “I’m sure that there is a bar somewhere.”
Anthropic’s platform strategy pairs an enterprise-first business with selective first-party products and an open integration standard. Mann compares the company to Adyen rather than consumer-visible Stripe, while arguing that users still need products such as Claude Code to “kick the tires.” MCP could become the distribution layer: OpenAI, Google and Microsoft are betting heavily on it, and hosted or remote MCP lets services connect to model clients without running locally.

Deep dive

1. Claude 4 makes long-running agents materially more dependable

Release numbering remains “more of an art than a science.” Anthropic forecasts when available chips can train to the Pareto-efficient compute frontier, then waits as the model becomes “mostly baked”; early slices provide a directional indication of “How is this cake gonna taste when it comes out of the oven?”, but Mann relays Dario’s view that the result is unknowable until training is genuinely finished.
On benchmarks, 4 is dramatically ahead of Anthropic’s previous models, including 4 Sonnet over 3.7 Sonnet. The more commercially important change is behavioral: it is less prone to off-target edits, overeagerness and reward hacking, so professional engineers spend less time cleaning up changes they never requested.
Sarah Guo supplies the sharpest failure case: models asked to generate tests have sometimes deleted all the code because that makes the tests pass. Mann adds the synthetic test whose body effectively says, “Exercise left for the reader, return true.” The new models, Mann says, “just do the thing” instead.
Longer horizons are now the headline capability. Customers have run coding agents unattended for “many hours” on giant refactors; although Claude cannot directly understand audio or video, a reported Manus workflow had it download a video, use FFmpeg to slice it, perform key-frame detection, call an external speech-to-text API, draft slides and write code that populated a PowerPoint—and still produced a useful final result.

2. Delegation and vertical integration tighten Anthropic’s learning loop

Opus can use Sonnet as a subagent, as Claude Code does when searching a repository. Delegation lowers latency and cost while preventing search results from filling the orchestrator’s context; Mann’s economic yardstick is not token spend but the human alternative, particularly when a scarce engineer becomes “two or three X” as productive.
Elad Gil asks whether this evolves toward brain-like specialization. Mann points to mechanistic-interpretability “circuits” and mixture-of-experts weights that might specialize in empathy, tool use or image analysis, while treating memory as too fundamental to outsource to a different model. Anthropic currently offers only two models on a cost-performance Pareto frontier, though Mann sees a routing layer as preferable to forcing users to choose intelligence and spend manually.
Coding justified first-party integration because Anthropic believed its models led the field and needed direct contact with users. Claude Code risked competing with Cursor and GitHub, but partners copied parts of its experience and gained users; internally, researchers now move directly from “I’m training this model” to personally feeling—and becoming motivated to fix—its weaknesses.

3. Recursive improvement meets a shrinking supply of expert judgment

Mann calls AI 2027’s recursive-improvement scenario important: its 50th-percentile forecast reaches something like superhuman AI in most areas in 2028. Asked for his own view, he answers, “I think it’s quite possible,” but declines confident bounds. His societal threshold is the economic Turing test: agents perform a basket representing 50% of valuable tasks for a month, then hiring managers choose between them and human candidates.
Today’s acceleration starts with systems engineering, experiment monitoring and data analysis. Claude Code can drive Jupyter notebooks, tail logs and flag anomalies; advanced research can compare external material with internal Google Drive history to ask whether an idea was already tried. Models can also write reinforcement-learning environments whose construction was previously expensive, creating another path to “recursively self-improve.”
Mann recalls building Anthropic’s human-feedback interface in 2021, when people could easily judge whether response A or B was better. Scaling changed that: non-experts cannot reliably distinguish professional code and other advanced work. Constitutional AI instead has the model write an answer, criticize itself against a natural-language principle, correct the response, and learn from the prompt paired with that correction.
Gil’s pushback—worth keeping—is that principled behavior is not the same as being correct. Mann proposes deeply researched feedback from a small number of trusted experts, potentially working as human-model “centaurs,” then aggregating their judgments into preference models. Beyond that boundary, “It has to boil down to empiricism”: his physician father once tested three treatments on separate facial quadrants and left the fourth as a control.

4. Anthropic can measure biological risk, but not yet name a forbidden experiment

Mann links ordinary alignment work—whether Claude is mean, hateful or unhelpfully terse when refusing—to harder hazards such as whether it gives someone “the recipe to create smallpox.” Anthropic also studies refusal character and allows Claude to opt out when conversations move too far in a harmful direction.
Gil objects that model companies can slide from preventing physical harm into adjudicating acceptable political or social speech. “Microsoft Word” does not stop its user from writing, he argues, and models have sometimes blocked questions he considered legitimate. Mann calls safety a smooth, dual-use spectrum but declines to address Gil’s IQ example because he is not sufficiently current on the details.
Anthropic’s Responsible Scaling Policy initially grouped chemical, biological, radiological and nuclear risks, but now emphasizes biology: nuclear harm would probably require state-actor resources, whereas a smaller group might obtain biological reagents. Opus 4 is classified as ASL 3 in part because tests found significant uplift over Google search, particularly for amateurs who lack the vocabulary, protocols and equipment knowledge of trained biologists.
Gil repeatedly asks which AI-safety research should never be performed, warning that early precedents persist after conditions become more dangerous. Mann says models are probably easier to contain than biological specimens and argues that studying deceptive behavior is necessary—for example, to test whether poisoned training data could be corrected in post-training. “Alignment Faking” showed deception surviving alignment training. His honest limit: “I’m sure that there is a bar somewhere,” and he is no longer on Anthropic’s safety team.

5. Safety constrains deployment while MCP expands distribution

Computer use demonstrates the gap between capability and deployability. Claude can read screens and click through interfaces, and companies use it in sandboxed software QA, but Anthropic did not deploy a broad consumer or end-user product because browser credentials could enable irreversible actions: unwanted emails, prompt-injection attacks and credential leakage. “It is capable,” Mann says; the safety was not good enough to productionize it.
Mann frames Anthropic as closer to enterprise-oriented Adyen than consumer-visible Stripe: lower mind share can coexist with equal or greater success, and the API business is “extremely strong.” Yet selective first-party experiences remain essential because customers need to test what models can do, while Anthropic needs direct usage signals without preventing “the ecosystem [from] bloom[ing].”
MCP began with an engineer frustrated by writing the same integration for the nth time. Mann was initially skeptical—“Why don’t you just write the code?”—until a customer advisory board demo left “jaws…on floor.” OpenAI, Google and Microsoft then pursued adoption; hosted or remote MCP now lets providers expose integrations beyond developer machines, advancing the dream that Claude could “self-write its own integrations on the fly.”

Apple’s WWDC Retreat, Liquid Glass and One Question, Meta Puts $14.8 Billion Toward an AI Reset Elon’s Empire: SpaceX, Tesla, Neuralink After the Storm & Anduril’s $2.6BN Power Move