Gemini 3.8 Flash

229 bratao 2026-09-02 15:21:28 UTC deepmind.google Llama Share 114

Gemini 3.8 Flash is a model that provides general-purpose, production-ready agents for cost-effective scaling of various tasks and workflows. It has limitations such as hallucinations and may exhibit slowness or timeout issues, but its safety performance is generally similar to Gemini 3.7 Flash.

A Note from LWN

440 rwky 2026-09-02 13:21:28 UTC lwn.net Llama Share 93

LWN is increasing subscription prices by 20% due to rising costs and inflation, effective September 15. The price increase will help maintain the site's independence from advertising and support its continued growth.

GrapheneOS says Pixel 11 has MTE support after all

90 user_7832 2026-09-02 14:21:28 UTC grapheneos.social Full Share 57

We have good news about the Pixel 11. It still has at least bare minimum support for MTE at a hardware level. We think they removed most of the hardware acceleration from the CPU cache to save money. They ruined the performance so it ended up being fully disabled in firmware. It may still be usable.

Biggest dark matter detector spots a single weird particle

109 randycupertino 2026-09-02 14:21:28 UTC science.org Full Share 20

Mistral now trains on user input by default, except on enterprise tier

171 teekert 2026-09-02 13:21:28 UTC help.mistral.ai Llama Share 79

Users can opt out of Mistral's model training programs at any time, with Vibe users opting in by default and Enterprise customers opted out by default. Users can disable the toggle to prevent data sharing for both Vibe and API services.

Three sites made 215,128 “best software” pages for AI. Perplexity cites them

140 jakobgreenfeld 2026-09-02 14:21:28 UTC trellner.com Llama Share 77

Two web-grounded models were asked to provide recommendations in 380 software categories, with 7,534 citations retrieved from their URLs. The majority of these citations point at domains ranked lower than #100,000 or not in the top million.

Gemini 3.8 Flash and 3.8 Flash Cyber

23 simonsarris 2026-09-02 15:45:28 UTC blog.google Llama Share 2

Gemini 3.8 introduces two variants, Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, offering significant improvements in software engineering, agentic tasks, and critical reasoning at the same speed and low cost as version 3.7 Flash.

Exit the Cave

56 akkartik 2026-09-02 14:21:28 UTC turtlespace.blog Llama Share 13

The author reflects on the "Cave" of private training and how it can lead to a false sense of progress, while missing out on real-world competition and growth. They share their personal experience as a wrestler, where they learned that true victory comes from embracing challenges and risks in public, rather than hiding in solitude.

Poisson Disk Sampling

55 vismit2000 2026-09-02 14:21:28 UTC stripeacross.com Llama Share 6

A mathematician discusses Bridson's algorithm for generating Poisson disk distributions, which efficiently places points in a grid with a minimum distance between them. The algorithm has been improved upon by adding parental optimization and changing the distribution of distances to generate more random configurations.

Aging Brains Blend Memories Together Instead of Just Forgetting Them

71 mdp2021 2026-09-02 13:21:28 UTC studyfinds.com Llama Share 26

Older adults' brains may blend memories together, causing mix-up errors, rather than simply losing memory accuracy. This blending effect is linked to stronger similarity in brain patterns between learning and recall phases.

Commodore 64 released September 1, 1982

240 giuliomagnifico 2026-09-02 09:21:28 UTC dfarq.homeip.net Llama Share 121

The Commodore 64, launched in September 1982, was the best-selling computer of all time with over 12 million units sold. It had a significant impact on its users' lives and careers, including the author's own experience with programming and IT.

A Selection of Los Alamos Rolodex Business Cards

40 1970-01-01 2026-09-02 14:21:28 UTC clui.org Full Share 7

A Selection of Los Alamos Rolodex Business Cards

WebLLM: high-performance in-browser LLM inference engine

27 saikatsg 2026-09-02 14:21:28 UTC github.com/mlc-ai Llama Share 7

WebLLM is a high-performance in-browser LLM inference engine that leverages WebGPU for hardware acceleration, enabling powerful LLM operations directly within web browsers. It supports OpenAI API and various models, including custom models in MLC format.

A third of Perplexity's citations don't contain the number they're cited for

46 jakobgreenfeld 2026-09-02 14:21:28 UTC hausresearch.com Llama Share 12

Perplexity's search models failed to verify 34.7% of citations, with 14.4% failing due to claims not being scored per citation rather than per claim. The models cited a significant number of gated URLs that could not be verified by readers without an account.

Six curl CVEs after OpenAI and Anthropic came back with zero

94 goobreee 2026-09-02 14:21:28 UTC aisle.com Full Share 30

Check if a file was made with Claude

73 frexs 2026-09-02 13:21:28 UTC claude.com Llama Share 48

This tool checks files for signs of Claude's involvement in creating or editing them, providing a C2PA content credential to verify its origin. A result indicates whether the content was processed by Claude, without revealing its substance or authenticity.

Show HN: FrontierHarness Eval – 9 harness, same model, cost per pass varies 17x

9 shiqimei 2026-09-02 16:16:28 UTC frontierharness.org LocalLLM Share 0

OpenCode's evaluation excludes failures, showing 15 passes and $3.24 per task when counting failed attempts. FrontierHarness v1.0 is designed for software engineering and terminal tasks, with a fixed environment setup for consistent testing.

LLMs: Intelligence vs. Cost

40 theanonymousone 2026-09-02 13:21:28 UTC openteams.com Llama Share 17

ArtificialAnalysis benchmarks LLM models' intelligence and costs, providing a useful Intelligence Index and cost comparison plots. The plots show diminishing returns on performance as cost increases, with cheaper models often sufficient for most users' needs.

AI, ex-Soviet engineers, and the Holy Grail of rocketry

8 rbanffy 2026-09-02 15:21:28 UTC fastcompany.com Full Share 0

Please enable JS and disable any ad blocker

Saving money on Google Photos with Immich: Your own personal photo storage

5 speckx 2026-09-02 16:09:28 UTC markpitblado.me LocalLLM Share 2

Hacker News discusses alternatives to paid cloud storage like Google Photos. Using Docker and Immich, users can set up a self-hosted photo server for free.

Telli (YC F24) is hiring engineers and designers [Berlin, on-site]

0 2026-09-02 09:21:28 UTC careers.telli.com Llama Share

Tellie is building AI to run customer-facing operations with agents that act, understand, and improve. The company prioritizes ownership, shared responsibility, conviction over compromise and agency over complacency in a merit-driven environment.

The Emergent Symbolic Structure of Artificial Neural Networks

240 schmuhblaster 2026-09-02 04:21:28 UTC arxiv.org Llama Share 83

Modern AI systems excel in domains they seem poorly suited for, despite being based on continuous vector representations. Researchers propose that these vectors implicitly realize symbolic structure, allowing them to achieve impressive performance in cognitive domains.

Dutch central bank moves share of gold from U.S., Canada to London

161 TechTechTech 2026-09-02 13:21:28 UTC nltimes.nl Llama Share 144

The Nederlandsche Bank has moved 86 tons of gold from US and Canada to London, increasing its share in the city. This move aims to improve crisis readiness by spreading risk and making Dutch gold more marketable.

9000 RPM is faster than your screen refreshes

35 Mojah 2026-09-02 13:21:28 UTC ma.ttias.be Llama Share 20

The user discusses the high speed of a car's crankshaft, comparing it to display frame rates and suggesting that capturing its motion at 150fps would reveal its true movement. The post also mentions Oh Dear, a monitoring platform for tracking system uptime and issues.

Ending my elixir exploratory writing

57 ffin 2026-09-02 12:21:28 UTC lucassifoni.info Llama Share 36

The author started a blog 3.5 years ago to share Elixir experiments and optics knowledge, which helped them attend conferences and meet friends. They now question the value of human creativity in programming communities dominated by LLMs.

The Cables That Connect the World

33 surprisetalk 2026-09-02 13:21:28 UTC xn--gckvb8fzb.com Llama Share 6

The internet infrastructure is a network of fibre-optic cables laid across the ocean floor, owned by various companies including Google and Meta. These cables carry data between continents and are often buried beneath the seabed to protect them from damage.

Using jq to format JSON on the clipboard

38 stefanvdw1 2026-09-02 13:21:28 UTC chris48s.github.io Llama Share 13

The user found a useful jq alias to format JSON in the clipboard, but needed to add error handling for non-JSON text. They created an updated alias with error handling using xclip and jq commands.

Move in C++ without a std:move

33 dalvrosa 2026-09-02 12:21:28 UTC andreasfertig.com Llama Share 39

The author advises using std::move rarely, citing return value optimization (RVO) and copy elision as the preferred methods to avoid unnecessary copies and moves. With C++23, implicit move is now supported in both cases where it was previously required or not possible.

Banca Etica Suspends A/I's Account While Condemning the Sanctions Behind It

54 rendx 2026-09-02 10:21:28 UTC sabot.media Full Share 33

It's OK to hardcode feature flags (2025)

60 biscuits1 2026-09-02 11:21:28 UTC code.mendhak.com Llama Share 35

Feature flags can be overhyped and introduce unnecessary complexity, security risks, and technical debt. Hardcoding feature flags in a simple JSON file is often the best approach, providing reliability and safety without the added overhead of management software.