Feed aggregator

Article URL: https://electrek.co/2026/03/05/floating-wind-turbines-could-soon-power-ai-data-centers-at-sea/

Comments URL: https://news.ycombinator.com/item?id=47271509

Points: 1

# Comments: 0

Article URL: https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-gpt-5-4-in-microsoft-foundry/4499785

Comments URL: https://news.ycombinator.com/item?id=47271502

Points: 1

# Comments: 0

Article URL: https://yadin.com/notes/generated/

Comments URL: https://news.ycombinator.com/item?id=47271481

Points: 1

# Comments: 0

I’ve been building ABES (Adaptive Belief Ecology System), a memory architecture for AI agents based on the idea that memory should manage belief state over time, not only retrieve prior text.

The system models memory as structured beliefs with explicit state, including confidence, salience, contradiction pressure, lifecycle status, memory tier, evidence balance, lineage, user and session scope, and decay behavior. Beliefs can be reinforced, weakened, contested, updated, mutated, or deprecated as new evidence arrives.

The goal is to support longer-running agents that need to deal with stale information, conflicting information, confidence change, and belief revision, rather than only recalling similar prior content.

Current implementation includes a structured belief model, reinforcement and decay mechanics, contradiction handling, tiered memory behavior, session isolation, API support, Docker support, and testing/evaluation infrastructure.

What has been verified so far in the project’s published tests and evals:

822 passing tests

a 1,000-prompt evaluation with an overall score of 825/1000 (82.5%)

reported category scores of 96.8% episodic memory, 94.4% working memory, and 92.8% semantic memory

a 15-block side-by-side evaluation against a raw Ollama baseline, where ABES passed 14/15 blocks and the baseline passed 6/15

a 200-prompt cognitive stress test reported as 3 consecutive runs at 200/200

Two easy verification points:

run PYTHONPATH=$PWD pytest tests/ -q from the repo root

inspect results/side_by_side_eval.json for the block-level comparison output

I do not consider internal tests and project-published evals to be sufficient external validation. The next stages are stronger benchmarking, improved contradiction handling and belief revision, stronger temporal and relational structure, longer-horizon testing, multi-agent shared memory work, and better observability of belief transitions.

Comments URL: https://news.ycombinator.com/item?id=47271473

Points: 1

# Comments: 0

Article URL: https://workshop.humancompatible.ai/#call-for-posters

Comments URL: https://news.ycombinator.com/item?id=47271426

Points: 1

# Comments: 1

Article URL: https://www.ft.com/content/162076f9-3eed-4b11-9a36-d5717c8b357b

Comments URL: https://news.ycombinator.com/item?id=47271416

Points: 1

# Comments: 0

I built this as a side project because I found existing family tree tools either too bloated or too manual—I spent more time dragging boxes and aligning lines than actually mapping my family history.

My goal was to make it instant: you just add the names, and the auto-layout engine handles the hierarchy and spacing automatically. It runs entirely in the browser and exports high-res PNG/JPGs.

Would love feedback on the layout logic (especially for complex families) and the overall UI flow. Happy to answer any questions!

Comments URL: https://news.ycombinator.com/item?id=47271396

Points: 2

# Comments: 0

I have recently asked the same question in a HN thread, which was mysteriously downvoted. The question remains to me: there is a lot of talk between Anthropic and the DOW about adopting LLM technology for warfare. Specifically, for "fully autonomous weapons and mass domestic surveillance". Does anyone understand how these two goals can be achieved? LLMs don't seem to me the right tool for this. Autonomous weapons would require a much faster and much more reliable and deterministic AI. LLMs might be a better use for mass surveillance, but I am not really sure how they would cope with the massive amount of data and the limited context window (unless they use the data itself for training). RAGs might only mitigate the problem. Does anyone have some ideas?

Comments URL: https://news.ycombinator.com/item?id=47271391

Points: 1

# Comments: 2

Article URL: https://dorrit.pairsys.ai/

Comments URL: https://news.ycombinator.com/item?id=47271374

Points: 2

# Comments: 0

Hey HN, I've been talking to AI and health-tech founders lately, and something keeps coming up: patents feel like they were designed for a different world. By the time you spend $20k and wait 2–3 years, your startup has probably pivoted twice.

So I'm genuinely curious what people here are actually doing. Are you filing patents anyway? Keeping things as trade secrets? Publishing defensively? Or just not thinking about IP until there's funding on the table?

I threw together a really short survey (60 seconds, promise) to get some real data—no sales, just trying to understand what founders actually do: Form Link: https://forms.gle/8UAytkGNfge4GKrH8

If you'd rather just comment below, that's honestly just as helpful.

Comments URL: https://news.ycombinator.com/item?id=47271361

Points: 1

# Comments: 1

Built this after one too many production incidents caused by a renamed variable or missing key that CI never caught. EnvSentinel treats your .env as a versioned contract - you define the rules once in a JSON schema, then validate against it in CI, pre-commit, or locally. Also regenerates .env.example from the contract automatically so it never drifts. Pure Python stdlib, no external dependencies, 3.10+.

Comments URL: https://news.ycombinator.com/item?id=47271348

Points: 1

# Comments: 0

Article URL: https://www.economist.com/science-and-technology/2026/03/04/a-basket-of-new-fruit-varieties-is-coming-your-way

Comments URL: https://news.ycombinator.com/item?id=47271338

Points: 1

# Comments: 0

Article URL: https://link.aps.org/doi/10.1103/2dyh-yhrb

Comments URL: https://news.ycombinator.com/item?id=47271337

Points: 1

# Comments: 0

Article URL: https://qcut.app/

Comments URL: https://news.ycombinator.com/item?id=47271311

Points: 3

# Comments: 2

Article URL: https://github.com/swyxio/claude-compaction-viewer/

Comments URL: https://news.ycombinator.com/item?id=47271295

Points: 1

# Comments: 0

Article URL: https://huggingface.co/blog/modular-diffusers

Comments URL: https://news.ycombinator.com/item?id=47271292

Points: 3

# Comments: 0

I often work on 2-3 projects at once. They each have some combination of: - terminal windows - browser for testing - browser for research, brainstorming, etc - documents & finder windows - various tools (expo, etc)

I have a lot of trouble keeping these separate. I use MacOS - you can have many desktops, but they're all in the same workspace. I think what I was is something like tmux for my whole computer, where I can switch away from a project and come back and be where I left off, with only the content from that project.

I actually tried to build this myself as the OS level, but Mac seems to lock everything down pretty hard.

Anybody have a good solution?

Comments URL: https://news.ycombinator.com/item?id=47271254

Points: 1

# Comments: 0

Article URL: https://www.tug.org/texlive/acquire.html

Comments URL: https://news.ycombinator.com/item?id=47271187

Points: 12

# Comments: 3

Hey HN, I built triplecheck because I wanted deep AI code review without paying $24/mo per seat.

The idea: instead of one LLM pass that drops comments (like CodeRabbit/Sourcery), triplecheck runs a full loop:

1. Reviewer finds bugs → structured findings with file, line, severity 2. Coder writes actual patches (search/replace diffs, not suggestions) 3. Tests run automatically to catch regressions 4. Loop until no new findings or max rounds 5. Judge scores the final result 0–10

The key insight: with local LLMs, compute is free, so you can afford to be thorough. Run 5 review passes from different angles, vote to filter noise, let the coder fix everything, and re-review until clean. Try doing that with a $0.03/1K token API.

What works well: - Qwen3-Coder on vLLM/Ollama handles reviewer + coder surprisingly well - Multi-pass voting genuinely reduces false positives — 3 passes agreeing > 1 pass guessing - Tree-sitter dependency graph means the reviewer sees related files together, not random batches - Scanned a 136K-line Go codebase (70 modules) — found real bugs, not just style nits

What's missing (honest): - No GitHub PR integration yet (CLI only — you run it, read the report). This is the #1 gap vs CodeRabbit. It's on the roadmap. - No incremental/diff-only review — it reviews whole files. Fine for local LLMs (free), wasteful for cloud APIs. - Local LLMs still hallucinate fixes sometimes. The test gate catches most of it, but you should review the diff before merging.

Stack: Python, Click CLI, any OpenAI-compatible backend. Works with vLLM, Ollama, LM Studio, DeepSeek, OpenRouter, Claude CLI. Mix and match — e.g. local Qwen running on M3 Ultra for reviewer/coder + Claude for judge.

Would love feedback, especially from anyone running local models for dev tools. What review capabilities would make you actually use this in your workflow?

Comments URL: https://news.ycombinator.com/item?id=47271100

Points: 1

# Comments: 0

On January 21, Google quietly changed the gemini-flash-latest alias to point to gemini-3-flash-preview — a model that does not support Google Search grounding.

The API never returned an error. HTTP 200, valid JSON, finish reason: STOP. The only thing missing was groundingMetadata. No warning, no deprecation notice, nothing.

I run PartsplanAI, a B2B electronic components marketplace. Grounding is not optional for us — we use it to verify part specs against real datasheets and prevent hallucination. Wrong capacitance values or voltage ratings from a language model aren't just embarrassing; they cause real problems for engineers downstream.

For approximately one month (late January through February 27), our AI features ran without any grounding. Every part recommendation, every spec search — generated purely from the model's pre-trained knowledge. The corrupted data accumulated in our database and was served to B2B customers. We had no idea.

On February 27, I noticed recommendations weren't matching real datasheets. What followed was 16 hours of debugging — 63 Git commits, 13 different approaches. I rewrote prompts, rebuilt the search pipeline, changed configurations, adjusted timeouts, switched between parallel and sequential calls. Nothing worked, because the problem was never in my code.

The fix: switch to gemini-2.5-flash. 20 minutes. Done.

The changelog entry for January 21 reads only: "gemini-flash-latest alias now points to gemini-3-flash-preview"

No mention of grounding regression. No compatibility warning.

There's also a second undocumented behavior: on gemini-2.5-flash, if you set responseMimeType: 'application/json' and googleSearch simultaneously, the search is silently ignored. No error, no docs, no warning.

GitHub Issue #384 (google-gemini/generative-ai-js) confirms the grounding issue was known in the community before the alias change was made.

The January 21 changelog was published the same week as the Gemini 2.0 Flash general availability announcement.

If you're using gemini-flash-latest with grounding, verify that groundingMetadata is actually present in your responses. You may have been affected since January 21.

Comments URL: https://news.ycombinator.com/item?id=47271099

Points: 2

# Comments: 0