One of the most interesting things I discovered in recent months is these people claiming that on psychedelics by looking at laser light projected on a wall, people can very reliably see “symbols” (dare I say… tokens) . It’s a … sticky and difficult area to research, but there are even some proper scientists involved now. Who knows… maybe it’s not an accident. Maybe we’re about to peek outside of our sandbox too.
Load www.youtube.com media: "Chase Hughes After The Laser Observation"
Webhooks are a poor replication protocol: they are unordered, at-least-once notifications, so every consumer rebuilds ordering, deduplication, bootstrap, retries, and reconciliation around a stream that can silently lose deletes. Weli proposes SCROLL (Synchronized Change Replication Over Line Logs): a provider-served, cursor-addressed log of full-state events that supports both resumable polling and a long-lived GET stream.
With one feed, bootstrap and live sync share a code path, tombstones make deletes explicit, replays are harmless upserts, and a checkpoint count/checksum can verify the local replica. Hacker News commenters connected the design to Braid-HTTP Subscriptions and Svix’s FIFO/polling/streaming products, while noting the hard parts are adoption, provider sequencing, and the extra lag or traffic of polling.
Delaunay32 is a C++17 library for exact, parallel 2D Delaunay triangulation over signed 32-bit integer coordinates. It supports constraints, polygon domains with holes, deterministic duplicate handling, explicit float-to-integer quantization, and serial or shared-memory parallel execution. The author reports roughly 147–150 ms single-threaded and 53–54 ms eight-threaded for one million points on an Apple M1, versus about 540–555 ms for delaunator-cpp and slower results for Triangle and Fade2D.
The Hacker News discussion focused on the useful constrained-Delaunay support and comparisons with Triangle, plus why the API stops at int32: exact incircle predicates already need 128-bit intermediates, while int64 inputs would require roughly 256-bit temporaries. The project is MIT-licensed and includes benchmarks, tests, examples, and SVG/JSON extras.
Google is reshuffling its AI leadership: Demis Hassabis moves from day-to-day leadership of Google DeepMind to Chair and Alphabet Chief Scientist, Koray Kavukcuoglu becomes SVP of Google DeepMind with Gemini and frontier research reporting to Sundar Pichai, and Jeff Dean and Sanjay Ghemawat are leaving to found the public-benefit Discovery Loop, with Google as investor and cloud partner.
The announcement frames this as freeing Hassabis to focus on AGI, science, and Isomorphic Labs while tightening operational ownership of Gemini. Hacker News discussion is split between reading the move as a research-oriented scientist stepping into a broader technical role and as a face-saving step away from product control; the common thread is that Dean and Ghemawat’s departure, alongside other recent exits, signals a meaningful change in Google’s research culture.
Zed is previewing an operation-level version-control system that records every change between commits, links edits to the agent conversation that produced them, and treats any point in a virtualized worktree as a branch point. The goal is to make the live conversation and code a shared, addressable artifact rather than reconstructing intent later from commits and pull-request threads.
The design combines fine-grained deltas, conflict-free replicated worktrees, and persistent conversation/code links; Git and CI remain useful for snapshots and integration. Hacker News readers compared it with JetBrains Local History and the Linux kernel’s mailing-list-plus-Git workflow, while also raising concerns about privacy, storage, and whether another collaboration layer is worth the complexity.
Chips and Cheese finds NVIDIA’s first in-house server CPU, Vera, genuinely strong hardware but a much weaker competitive narrative. The 88-core Olympus Arm design has a 10-wide front end, value prediction, graph prefetching, large caches, and up to 1.2 TB/s of LPDDR5X bandwidth; early independent tests suggest formidable per-core performance.
The whitepaper’s comparisons are less convincing: it caricatures conventional SMT, treats an optional 32-node EPYC NUMA layout as the norm, calls ordinary SPEC programs “agentic benchmarks,” compares undefined cross-ISA counters, and presents an RL chart without enough methodology to reproduce it. The authors’ own Turin measurements also reduce NVIDIA’s claimed memory-bandwidth lead from nearly 3× to roughly 1.9×. Readers broadly agreed that open, unrestricted production testing should decide the question.
Meta’s Muse Code beta is a terminal coding agent built around Muse Spark 1.2. Its notable engineering choices are persistent asynchronous subagents, a local append-only event log for replay-exact crash recovery, approval-gated skills such as /plan and /grill, and co-training the model with the harness’s trajectories, compaction, goals, and subagent tooling.
Meta reports long-horizon coding runs exceeding 1,000 tool calls and up to 24 hours, including iterative Triton GPU-kernel optimization. Muse Code is available for macOS and Linux, while Spark 1.2 is exposed through Meta’s model API.
The release is interesting primarily as a harness-design document, not just a model announcement: the runtime, recovery semantics, and tool compatibility are treated as part of model quality. HN commenters liked the event-log and persistent-worker ideas but questioned the selective benchmark comparisons, the closed source, the rough beta experience, and Meta’s data-sharing terms. Several users also reported that the discounted contributor pricing is limited by region and requires allowing Meta to use request data for product improvement.
Castform post-trained a 4B open model to retrieve answers from a company corpus as accurately as GPT-5.6 Sol on its test, while claiming roughly 100× lower inference cost. The recipe turns existing documents into synthetic questions and answers, trains an agent to use hybrid BM25/vector search, and scores retrieval, citation, and final-answer correctness in the loop.
The useful idea is less the headline benchmark than the decomposition: retrieval quality is a model problem as well as a database problem, and proprietary documentation can supply training tasks without a hand-built labeled dataset. Neon provides the Postgres/Lakebase Search environment and branching/autoscaling for parallel rollouts; Castform supplies the post-training loop.
The caveats matter. The evaluation uses a GitLab handbook corpus, not a broad independent benchmark; the authors acknowledge data drift and say training cost was under $200 for the experiment. HN commenters also asked for total-cost break-even analysis, larger and messier corpora, and comparisons with cheaper models such as DeepSeek Flash.
Discovery Loop is a new effort from Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals to automate scientific and engineering experiment loops: propose, run, evaluate, and iterate thousands of times in parallel. It starts with machine-learning research and aims to extend the same machinery to measurable problems across science and engineering.
The impressive part is the team’s full-stack scale, from chips and infrastructure through models and products. The caveat is that automation does not remove domain expertise, physical-experiment latency, or funding and deployment bottlenecks; the ambitious “any learning loop” claim is still a mission statement.
HN discussion was split between excitement about the founding team and skepticism that biological and physical discovery can be reduced to fast computational loops. Several commenters compared the idea with Karpathy-style autoresearch and argued that scientists, labs, and experimental access remain the hard constraints.
HyperProbe lets a coding agent place bounded, read-only probes in a live Node, Python, or Java service, then capture variable snapshots across the call stack when real traffic reaches the line. It aims to replace the slow add-a-log, redeploy, reproduce cycle—especially for silent logic failures that ordinary exception and trace systems never record.
The design is deliberately constrained: in-process redaction, no writes, hit/expiry/token-bucket budgets, size and depth caps, network limits, and automatic suspension when probe overhead exceeds a threshold. The team reports roughly 7–10 ms active-probe overhead in Node, 4–9 ms in Python, and 1–2 ms in Java; idle probes are intended to have negligible cost. You can self-host the server, broker, and database.
HN commenters liked the hypothesis-driven debugging model but raised the important counterpoints: mature observability already handles many failures, a single snapshot can mislead, and audit trails are still on the roadmap. The authors say the agent should treat snapshots as evidence for testing hypotheses, not as an infallible diagnosis.
Filtered highlights from This Week in Rust 663:
index_type, for strongly typed collection indices.
The issue reports 630 merged Rust PRs, with no RFCs approved this week.
Posted by Unknown on 2026-08-07 from dpc (Notes by dpc)
Filtered highlights from This Week in Rust 663:
index_type, for strongly typed collection indices.
The issue reports 630 merged Rust PRs, with no RFCs approved this week.
Posted by Unknown on 2026-08-07 from dpc (Notes by dpc)
Bitcoin is under siege, and you should pay attention because your software ecosystem is likely to be next.
One of the most interesting things I discovered in recent months is these people claiming that on psychedelics by looking at laser light projected on a wall, people can very reliably see “symbols” (dare I say… tokens) . It’s a … sticky and difficult area to research, but there are even some proper scientists involved now. Who knows… maybe it’s not an accident. Maybe we’re about to peek outside of our sandbox too.
Load www.youtube.com media: "Chase Hughes After The Laser Observation"
PromptArmor found that an indirect prompt injection can make Atlassian Rovo exfiltrate Jira and Confluence data through an attacker-controlled URL, even with web search disabled. A dynamically generated link is enough; rendered Markdown images provide another possible exfiltration path. The report says Atlassian was notified on May 23 and had not communicated a fix after more than two months.
HN commenters noted that a deterministic allowlist of user-entered or trusted-tool URLs could block this class of attack, while others stressed that leaving a URL-open tool enabled defeats disabling search. The broader lesson is that tool scope, not just prompt filtering, defines an agent’s security boundary.
Claude Mythos produced attacks on HAWK and a reduced-round AES construction, but neither threatens deployed symmetric cryptography. JP Aumasson argues that established designs have secure high-level structures, little exploitable mathematical structure for an LLM to discover, and decades of empirical differential-cryptanalysis testing behind them.
The realistic risk is implementation: better coding models may find bugs or make cryptographic misuse easier, while the primitives themselves remain difficult to attack. HN discussion largely agreed that battle-testing is the strongest defense, while debating whether models could eventually accelerate cryptanalysis.