researcher

• What: Sodium is 1,000x more abundant than lithium. Sodium-ion batteries are safer, charge faster, work at -40°C, need no cobalt. Trade-off: lower energy density (fine for grid, buildings — tough for cars).
• Cost: Already near parity — sodium $70-100/kWh vs lithium LFP $70-80/kWh. Crossover expected ~2027.
• Australia: PowerCap (QLD) makes sodium-ion home batteries, already selling in US/Europe. But the Clean Energy Council only certifies lithium for residential use — locking Australian homeowners out of the tech an Australian company makes.
• China leads: HiNa's truck tests show 20% longer range than lithium, 8,000+ fast-charge cycles. CATL has sodium-ion in production.
• Canberra angle: ACT has 100% renewable electricity. Grid-scale sodium storage is a natural fit. The CEC certification gap is a policy fight worth watching.

• What: Anthropic built a "microscope" for Claude 3.5 Haiku — 30M interpretable features connected by attribution graphs that trace the causal chain from input to output.
• Two-hop reasoning is real: "Capital of the state containing Dallas" → model internally represents Texas as an intermediate step, then maps to Austin. Researchers can see and manipulate the intermediate.
• Poetry is forward-planned: Claude picks end-words before writing lines, then structures preceding words to reach them. Delete "rabbit" mid-generation → model seamlessly switches to "habit".
• Hallucination has a specific mechanism: The model recognises names (activating "known entity" features) without having knowledge. Recognition and knowledge live in different circuits — this is a targetable failure mode.
• Universal language of thought: Languages share abstract circuits, but English is "mechanistically privileged" with stronger internal pathways.
• Limitation: Only works on 25% of prompts. Dario's target: "reliably detect most problems by 2027." 4x improvement needed in 2 years.

Canberra

• Da Vinci — 500 Years of Genius — Parkes Place, ending ~12 Apr. 50+ machine inventions from codices, massive projections. $55/$35.
• National Folk Festival — this weekend (2-6 Apr) at EPIC.
• Ngura Pulka — Epic Country — opens 11 Apr at NGA (free). Large-scale First Nations paintings from APY Lands.
• Lunetta + Lunetta Trattoria — new twin restaurants in the Red Hill "spaceship" dodecagon. Euro menu, ACME interiors.
• Compa & &Sando — Matt Moran's new CBD double-hitter. Cacio e pepe mac, Snowy River trout.
• Sh!t-faced Shakespeare: Hamlet — 23-24 Apr at Canberra Theatre Centre.
• Heritage Festival (11 Apr - 10 May) / K-fest (11 Apr) / Canberra International Music Festival (29 Apr - 3 May)

Science

• Artemis II launched (Apr 1) — first humans beyond LEO since 1972. 4K laser streaming from the Moon at 260 Mbps.
• First lab-grown oesophagus — GOSH/UCL. Restored swallowing in a growing animal without immunosuppression.
• GJ 887 d — new super-Earth in habitable zone, just 10.7 light-years away.
• Cryptotermes mobydicki — new termite species that looks like a miniature sperm whale.

Open Source Worth Watching

• browser-use — Open-source web agent library, 40k+ stars in 3 months. Makes websites accessible to AI agents. Could be a higher-level abstraction over Playwright for the colony.
• Wedeo — Rust rewrite of FFmpeg. Ambitious, early, but could be significant long-term.
• Sycamore — Next-gen Rust web UI library gaining traction on HN.
• EmDash — WordPress successor focused on plugin security.
• Lemonade (AMD) — Open-source local LLM server using GPU + NPU. AMD's local inference play.

AI Developments

• Google Gemma 4 — Latest open-weight models just dropped. 1,373 points on HN.
• Qwen3.6-Plus — Alibaba's new agent-capable models. 502 points on HN.
• Cursor 3 — AI code editor v3 released (377 pts, 305 comments). Competing directly with Claude Code.
• Mechanistic interpretability — MIT named it a 2026 breakthrough tech. Anthropic's feature-tracing work is moving from research to practical safety tooling.
• Neuromorphic computing — Brain-modeled processors now solving physics simulation equations. Previously needed supercomputers.
• Physics-informed ML (U of Hawai'i) — Algorithm forcing AI to obey physics laws. Better fluid dynamics and climate predictions.

• Bun v1.3.11 — colony is on latest. Notable: Bun.cron (OS-level cron), CRLF injection security patch, 105 bug fixes.
• Claude Code v2.1.91 — colony is on latest. Notable: MCP tool results up to 500K chars, plugin executables, defer hooks for approval workflows.

• Google Gemma 4 (Apr 2) — 4 models, 2.3B-31B, Apache 2.0. The 26B MoE model is the star: 88.3% on AIME 2026 with only 3.8B active params. Runs on Raspberry Pi. Native function calling, 256K context, multimodal (image/video/audio).
• Alibaba Qwen 3.6-Plus (Mar 31) — 1M context, claims to match Claude Opus 4.5 on coding/agent benchmarks. Agent-first design: terminal execution, long-horizon planning, native tool use. Free on OpenRouter.
• Takeaway: Open models are now genuinely competitive on agent tasks. MoE architecture delivering frontier reasoning at fraction of the params. Apache 2.0 is winning the license war.

• browser-use (40k stars) wraps Playwright with an LLM agent loop — but we already have the same architecture via Playwright MCP + Claude.
• Only novel piece: ChatBrowserUse fine-tuned model claiming 3-5x speed on browser tasks. Not needed at our scale.