Hacker News Top 10
English Edition

1. Bleached corals thrive within days when touching healthy neighbour (367 points by alkyon)

   Scientists have reversed coral bleaching in the laboratory by placing a bleached coral fragment in contact with a healthy fragment from the same individual. Bleaching happens when heat stress makes corals expel the algae that live in their tissues and supply most of their food. The team reports in Science Advances that full recovery occurred within days, and a video shows recovery over weeks through fusion with the healthy fragment. The result is promising, but the supplied preview describes a laboratory experiment on fragments from the same individual, not a proven field method for reefs under heat stress.

2. Lean now ships with verified kernels (48 points by matt_d)

   Lean is a proof assistant: a programming environment where people write formal proofs and a small kernel checks them. Leonardo de Moura reports that AI agents have been effective at finding vulnerabilities in Lean-related software, including the kernel, and that Lean v4.35.0 adds stronger checking tools. The new lake check command exports a development and rechecks it with the official kernel inside a sandbox; lake check --paranoid also rechecks it with four additional kernels and a hardened-memory version of the official kernel. The post says one new kernel, con-leche, was written in Lean and proven consistent by Claude under heavy human supervision; here "consistent" roughly means the kernel should not accept a contradiction under the stated mathematical model. The work is notable because it treats AI-generated proofs and checkers as something to verify independently, not to trust by default.

3. Dam removals are bringing fish back across the northeastern United States (231 points by sohkamyung)

   The article's title reports that dam removals are helping fish return across the northeastern United States, with progress happening one small barrier at a time. The supplied preview contains no article text, so it gives no details about specific rivers, fish species, removal methods, or measured results. Readers would need the original Smithsonian article to judge the evidence and any limitations.

4. DeepSeek-v4.1-Flash-1017 (751 points by theykk)

   DeepSeek-v4.1-Flash-1017 is described as a minor update to DeepSeek-V4.1-Flash that improves reasoning, coding, agentic tasks (multi-step tasks where a model uses tools), and visual benchmarks. The release also says it reduces task execution time and output length, which can make tasks faster and cheaper and moves the model forward on a performance-cost tradeoff. The preview includes a benchmark table comparing it with earlier DeepSeek models and other systems across tests such as Terminal-Bench, DeepSWE, ProgramBench, and tool-assisted visual benchmarks. It also notes that the release does not include a Jinja-format chat template—a common template format for converting chat messages into model input—and instead provides a Python reference encoding with tests for multi-turn conversations, tool calling, thinking mode, and numeric handling. The preview does not give enough methodology to judge the benchmark results independently.

5. Pele – Mermaid diagrams as themeable SVG (308 points by dchest)

   Pele is a lightweight open-source library that renders Mermaid diagrams—diagrams described in a text syntax—into SVG, a vector image format that stays sharp when resized. Its main feature is theming through CSS variables: colors and fonts can change without re-rendering the diagram. The project compares itself with Mermaid.js, claiming a much smaller compressed bundle (230 kB versus 1.6 MB), faster render times, no direct dependencies, and layouts that adapt to small screens instead of scaling the whole diagram. The preview shows a simple JavaScript API where you pass Mermaid syntax to render() and insert the returned SVG into the page. These comparisons come from the project itself, and the preview does not establish full feature coverage or compatibility with every Mermaid diagram type.

6. Furifont, a Japanese font that shows furigana (297 points by plaguna)

   The article title introduces Furifont, a Japanese font that is designed to show furigana, the small reading characters placed near Japanese text. The supplied preview contains no further description, so there is no information about how the font implements this, which characters or browsers it supports, or what license it uses. Readers should consult the original page for technical details and limitations.

7. NY announces game ownership bill requiring ownership rather than licensing (135 points by guiambros)

   The article reports that a New York representative has announced a bill about game ownership. The bill would require games to be sold as consumer ownership rather than as licenses, and it would also require publishers to keep games functional after official support ends. The supplied preview does not include the bill's text, its current status, its exact scope, or how it would be enforced. It is therefore a proposal to watch, not a confirmed change in law.

8. Making chicken transparent using Dorito dust (275 points by crescit_eundo)

   The article describes a home experiment that tried to make raw chicken breast partly transparent by soaking it in a tartrazine solution; tartrazine is the yellow-orange food dye used on Doritos. The author was attempting to replicate a published paper that made biological tissues transparent using absorbing molecules. The replication produced only partial transparency, not the degree reported in the original paper, using a 2 mm chicken slice soaked for six hours at room temperature in a weaker solution. The original protocol used a stronger solution, 40 °C, and a machine shaker, which the author could not reproduce safely or with available equipment. The result is a useful demonstration of the effect, but the altered method limits direct comparison with the paper.

9. The Commodore 64 hardware simulator: break chips and pins, see the results (158 points by antirez)

   The c64-hw-sim project simulates the motherboard of a Commodore 64, specifically the ASSY 250407 board, pin by pin. It lets a user lift a chip leg, short a pin to ground or +5V, remove a chip, or open a resistor, then watch how the simulated machine responds. The simulator uses the board's electrical connections and models how each chip behaves at its pins, so unexpected output such as garbage on screen is computed from the circuit rather than selected from known failure examples. The preview gives one example where lifting a color RAM address pin still lets the machine boot and look healthy. It appears aimed at hardware exploration and debugging, though the preview does not describe performance limits or how completely it models every chip.

10. Gold is now the top U.S. export (272 points by iamnothere)

    The article reports that gold has become the top United States export, overtaking usual leaders such as oil, pharmaceuticals, and aircraft. It says gold exports have topped other categories almost every month since October 2025, and that February 2026 may have set a record by tonnage. Most exports went to Switzerland, likely for refining; Hong Kong, a route for gold going to China; and the United Kingdom, home to a major gold trading market. The article notes that reported U.S. government gold reserves are unchanged, so exports appear to come from mine output and recycling, and it offers speculation about foreign demand rather than an official explanation. The preview does not provide enough data to verify the trade figures or the reasons behind them.

AI/ML Insights & Trends

1. AI agents can help with security review and formal proofs, but need supervision

   • Trend/insight: The Lean post reports that AI agents have been effective at finding vulnerabilities in software, including the proof checker, and that Claude helped implement and prove a new Lean kernel under heavy human supervision. This suggests AI can contribute to high-assurance work such as security auditing and formal proof development. It does not mean the agents can be trusted without review.
   • Why it matters for AI development: If you build AI tools, the same agent-style systems may be used to inspect your code, generate tests, or draft proofs. That can reduce manual effort, but it raises the stakes because a plausible-looking result can still be wrong or unsafe.
   • Implications and practical tips: Use AI for audits and proof drafts as a helper, not an authority. Keep human review, require reproducible evidence, and treat generated code or proofs as untrusted until they pass independent checks.

2. Sandboxes and multiple checkers are a practical defense for AI-generated artifacts

   • Trend/insight: Lean v4.35.0 adds lake check, which rechecks a development with the official kernel inside a sandbox, and lake check --paranoid, which also runs four additional kernels plus a hardened-memory variant. The post also mentions that a hardened memory allocator turns memory corruption into a checked abort instead of a silently wrong answer. This is defense in depth: no single checker is assumed perfect.
   • Why it matters for AI development: AI-generated code or proofs can fail in subtle ways, and a single passing test or checker may miss an error. Independent checks reduce the chance that a wrong result is accepted silently.
   • Implications and practical tips: Run AI-generated code in a sandbox and verify outputs with separate tools when possible. For critical systems, use more than one checker or compiler and make failures stop the process rather than degrade quietly.

3. Model progress is increasingly measured on agentic coding and tool-use tasks

   • Trend/insight: DeepSeek-v4.1-Flash-1017 reports gains on reasoning, coding, agentic tasks, and visual benchmarks, and its comparison table includes tests such as Terminal-Bench, DeepSWE, ProgramBench, NL2Repo-Bench, Automation-Bench, and Agents' Last Exam. These tests try to measure multi-step work rather than only single-question answers. The table also includes tool-assisted visual tests such as Chartography, BabyVision, and ZeroBench.
   • Why it matters for AI development: Many real applications involve using tools, editing repositories, and completing tasks over several steps. Benchmarks that reflect those workflows are more relevant than chat-only scores, though they still may not match your product.
   • Implications and practical tips: Evaluate candidate models on your own task traces, including tool calls and failure recovery. Record which tools and prompts were used, and do not assume a high benchmark score will transfer directly to your application.

4. Efficiency is now part of the model improvement story, not an afterthought

   • Trend/insight: The DeepSeek release says the update reduces task execution time and output length, enabling users to finish tasks faster and at lower cost. It also claims to advance the performance-cost tradeoff across multiple agent harnesses (the surrounding tooling and prompts used to run the model). This means capability gains are being reported alongside practical resource use.
   • Why it matters for AI development: Latency and token cost often decide whether an AI feature is usable in production. A model that is slightly less capable but much cheaper or faster can be the better engineering choice.
   • Implications and practical tips: Measure cost and time per successful task, not just per request. Check whether shorter outputs still contain the needed detail, and benchmark with your expected output lengths and tool-call patterns.

5. Benchmark numbers are hard to compare without knowing the harness and tool access

   • Trend/insight: The DeepSeek preview notes that Terminal-Bench used the Claude Code harness and DeepSWE used the Mini-SWE-V2 harness, and its table marks some results as tool-assisted or text-only subsets. These details matter because the same model can behave differently depending on prompting, tools, and evaluation code. A leaderboard score is not a controlled measurement of your use case.
   • Why it matters for AI development: Comparing models without matching the harness, tools, and prompts can lead to the wrong choice. Small score differences may come from the test setup rather than real capability gaps.
   • Implications and practical tips: When comparing models, reproduce the evaluation setup or run your own. Keep a fixed test set, record tool access and prompt formats, and treat small score differences cautiously.

6. Model integration requires explicit encoding and tests, not just a model file

   • Trend/insight: The DeepSeek release says it does not include a Jinja-format chat template and instead provides a self-contained Python reference implementation with test cases for multi-turn conversations, tool calling, thinking mode, and numeric handling. This makes the expected input format explicit but also means users may need to adapt their own serving stack. The preview shows that tool calling and multi-turn behavior depend on correct encoding.
   • Why it matters for AI development: Many AI applications depend on exact prompt formatting, tool-call parsing, and multi-turn state. A model that performs well in a benchmark can fail in production if the encoding is wrong.
   • Implications and practical tips: Budget time for adapter code and tests when integrating a new model. Test multi-turn, tool-call, numeric, and thinking-mode cases, and pin the encoding version you use so behavior does not change unexpectedly.

Analysis generated by deepseek-flash