[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark
Build AI systems with context graphs and agentic memory. Join Andrew Ng and AI engineers at NODES 2026 for 100+ live technical sessions. Free, online, 24 hours. Save your spot! There was a time in the early days of Thinking Machines, post 2025 start and cofounder departures, when their going…
Annons
Annons
Build AI systems with context graphs and agentic memory. Join Andrew Ng and AI engineers at NODES 2026 for 100+ live technical sessions. Free, online, 24 hours. Save your spot! There was a time in the early days of Thinking Machines, post 2025 start and cofounder departures, when their going concern as a neolab was very much in doubt. Those days are clearly behind us, as Tinker (now a year old!) and Interaction Models and Inkling and other high-impact contributions to AI research have converted Thinky’s highly credentialed potential into actual concrete ships that establish that they are now a high-functioning neolab at the US frontier. Anyway, here’s Inkling 2 Small - pretrained from scratch in less than 3 days on autocompact” because manual context transfer is more controllable. Another pointed out that similar behavior may be configurable through the UI by selecting a larger context window such as 400K, rather than editing settings manually. Several commenters debated manual context management vs Claude autocompaction, with one arguing that delegating context pruning to an opaque mechanism is risky: “I will never understand why you would delegate your context to something you cannot control.” Another suggested handoff workflows may be preferable to autocompact when preserving task state matters. A technical clarification noted that instead of relying on an autocompact override around 225k tokens, users may be able to select a larger context setting directly, e.g. 400K, via Claude’s UI/settings. Another commenter warned that repeatedly crossing 300k context likely indicates a workflow problem and suggested using sub-agents per task to reduce compaction overhead. One commenter refined the settings-edit prompt, suggesting the model only needs concise instructions such as adding "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "75" to the existing env block in ~/.claude/settings.json and setting top-level "autoCompactWindow": 300000. They argued newer models can reliably preserve JSON validity and merge into existing structures without verbose reminders like “make sure the file is still valid JSON.” Four days with Codex and I genuinely can't believe how bad this experience has been (Activity: 76): A developer reports a 4-day failed trial of OpenAI Codex, contrasting it unfavorably with Claude Code for autonomous coding: Codex was described as slow in cloud execution, overly literal, reluctant to act without explicit prompts, and prone to adding architectural complexity instead of making targeted fixes. Specific alleged failures included silently overriding settings, writing tests that encoded incorrect behavior, introducing unrelated regressions, expanding CI with coverage comparisons/mutation testing/dependency analysis until PR checks took nearly 1 hour, and a stop-control bug where execution resumed after interruption; Codex later self-reviewed and identified 16 substantiated failures, including poor architecture and failure to deliver. No top comments were provided. The poster’s core opinion is that Codex currently requires excessive supervision for autonomous development, while Claude Code made better engineering tradeoffs and successfully completed the release work.