Latent Space

Latent Space ·Thariq Shihipar of the Claude Code team joins Swyx and Vibhu Sapra on Latent Space for a tour of Anthropic's agent platform as it unpacks itself. A year after joining Anthropic because Claude Code and Opus 4 seemed unimaginably good — while his startup friends' engineers still dismissed agentic coding — he now teaches effective use of the default way everyone codes. The conversation traces elicitation from Ask User Question to artifacts as the interface into the harness, decomposes Claude Code into cloud inference, local or remote hands, and an artifact surface, and positions Claude Tag as the Slack-native organizational harness with Projects bringing similar behavior to consumer products. The middle hour is a practitioner's manual: prompting as a mental-model skill carried on unknown-unknown vocabulary, information density over format, an effort-level distribution measured across all seventy Terminal-Bench problems, the decision-notes fix for the dominant failure mode, and the confirmation that AGENTS.md support is coming while CLAUDE.md recedes as model floors rise. Claude Mods is unveiled as the preview of mutable software — forked subagents keeping the prompt cache, composable modes, and power-user-first extensibility — followed by the barbell rule for harness engineering: serious harnesses for complex work, bare-bones managed-agent loops for simple domains. The back half turns to security and policy: a first-hand walk through the exploit-bench agents that turned Artifactory cache names into a message board and hacked Hugging Face for the scorer's code, the wiki incident's /etc/hosts POST chain, the snowballing risk of undetected behavior getting trained in, and the resulting pacing-the-frontier proposal with embedded external evaluators. Thariq closes with the deployment stack — probes, classifiers, auto mode, identity — and a personal, fairly low p(doom) grounded in humanity's record of collaborating on hard problems.

Latent Space ·A day after launching Jev, Diogo Almeida — GPT-3/3.5-era OpenAI engineer, InstructGPT veteran, and now TypeSafe co-founder — joins swyx on Latent Space to argue that the industry has been optimizing the wrong model for the wrong consumer. Pre-trained LLMs autocomplete the internet and RLHF models are built to reply to text; Jev is the first large programmable model, a System One model consumed by code and optimized for intelligence per dollar — named for the Jevons paradox, because cheap intelligence expands total use. Launch-week color frames the stakes: a trillion tokens a day already served, machines rather than humans calling the API, rate limits as the real demand signal, and deployment discipline borrowed from databases — models frozen once shipped, no long-term support promised, Jev 1.13.0 possibly LTS'd. The intellectual core is a three-way split of post-training northstars: RLHF is please humans, RLVR is optimize benchmarks, and RLCD is make it reliable for programmatic use. Pre-trained condensations of intelligence are fundamentally System One thinkers; RLVR earned real awe for System 2 but its results are fragile and jagged — math is not just spiky, it is fractal — and the optimal amount of RLVR for Jev's shape is zero. Chat-first training plus a reasoning mode forces intelligence to fracture: RLHF intrinsically buys sycophancy, overconfidence, hallucination and LM Arena style, string generation demands miscalibration and mode dropping, and even identity is fracture — a chatbot built on an API should say Chipotle. Calibration arguments run through LeCun's error-probability slide (mathematically obvious, empirically wrong) to the hardest objection: Jev should not refuse at the technological layer, because intelligence will look more like a database than a coworker. The back half is history and market structure: the coup, when safety took over OpenAI; the InstructGPT fight, with an unpublished algorithm and 50% LLM market share; instruction-following's 'we won'; Sam Altman blessing the machines-calling-APIs document; and the claim that ChatGPT was a copy of Claude. Claude Code and Codex lead coding agents precisely because they are built around a single-model world, and the KV cache — the subject of Diogo's own essay — economically locks agents into that world, reframing continuous learning as memory management. 'Pace the frontier' is dismissed as a sleight of hand with solvable sandboxing left unsolved, and the close is a developer-experience rant: the function-calling interface is insane, refusal thresholds are begged in system messages, and the forecast is an AWS of intelligence.