AI News Today

I've been running Qwen3.6-35B-A3B locally on llama.cpp and noticed that prompt processing throughput gets too low with MTP. I got nerd-sniped. What started as curiosity turned into a two-week rabbit hole of experiments and ended with a PoC that fully recovers the MTP PP overhead on GPU, above any expectation I had. TL;DR: instead of processing the last layer MoE FFN for the entire ubatch tokens (usually 512-2048 tokens), this PoC processes only the output row (usually 1 token during prefill). Th

A radical solution. ScienceAlert stories are written, fact-checked, and edited by humans, never generated by AI. Don't miss a story, subscribe here.

Maraget Atwood, the storied author of The Handmaid's Tale and The Blind Assassin, was interviewed as part of the Babell Literary and Cultural Festival in Porto, Portugal. As it usually does at these things, the issue of AI came up, and Atwood didn't mince words. According to Deadline's recap, Atwood said she'd used an AI […]

Article URL: https://github.com/KayhanB21/riskratchet Comments URL: https://news.ycombinator.com/item?id=48700587 Points: 2 # Comments: 0

Recently saw this comment on another HN thread about the US government gating access to GPT-5.6 and how it harkens back to the 1990s encryption-as-export-controllable-tech situation and how people tattoo'd the algo to their bodies: > I can't wait for the first person to tattoo model weights on their body! (https://news.ycombinator.com/item?id=48693721) And I can't help but wonder, what would be the smallest functional amount of weights from any sort of ML model you could realistically tattoo ont

The lawsuit dates back to a deadly incident in Arizona in 2023.

Article URL: https://www.nippon.com/en/japan-topics/c030329/ Comments URL: https://news.ycombinator.com/item?id=48700324 Points: 3 # Comments: 0

<pre style='white-space:pre-wrap;width:81ex'>fix(model-fallback): don't rethrow provider-side AbortErrors as user cancellations (#90908) * fix(model-fallback): don't rethrow provider-side AbortErrors as user cancellations When the LLM API closes the connection mid-stream, the fetch layer surfaces AbortError("This operation was aborted") with no external abort signal triggered. The old guard `shouldRethrowAbort()` returned false for these errors (because isTimeoutError matched the message), so th

Article URL: https://www.bloomberg.com/news/articles/2026-06-26/chinese-hedge-funds-warn-the-ai-super-bubble-is-ready-to-burst Comments URL: https://news.ycombinator.com/item?id=48700240 Points: 7 # Comments: 1

Article URL: https://www.napster.com Comments URL: https://news.ycombinator.com/item?id=48700041 Points: 3 # Comments: 2

Article URL: https://www.hyoomn.com/?0 Comments URL: https://news.ycombinator.com/item?id=48699960 Points: 1 # Comments: 1

Article URL: https://www.bleepingcomputer.com/news/security/clean-github-repo-tricks-ai-coding-agents-into-running-malware/ Comments URL: https://news.ycombinator.com/item?id=48699886 Points: 4 # Comments: 0

Anthropic's AI model, Fable 5, could be available again within days. According to Axios, the Trump administration is close to lifting the restrictions imposed on June 12 over safety concerns. The Pentagon and NSA still need to sign off. The article Anthropic's Fable 5 could return within days as Trump administration prepares to lift restrictions appeared first on The Decoder .

The setup that makes Claude Code fit into your designer workflow Continue reading on UX Collective »

Paul Meade, the Apple vice president in charge of the Vision Pro headset, is reportedly leaving the company to join OpenAI’s hardware team.

<!-- SC_OFF --><div class="md"><p>Hey guys,</p> <p>I was playing around with Nanotron recently and got super frustrated by how many heavy, hardware-specific dependencies it imports at the module level ( flash-attn , triton, functorch , etc.). If you try to run it on older or budget GPUs like a T4 or V100, it just crashes on import.</p> <p>So I wrote Picotron (<a href="https://github.com/Syntropy-AI-Labs/picotron">https://github.com/Syntropy-AI-Labs/picotron</a>) to solve this. It's a clean-room

I’m from Pakistan and I currently work as a Backend Engineer remotely. But given my background in Electrical Engineering, I keep tinkering with electronics and robotics stuff whenever I have time to spare. Currently I’m working with Embedded Rust using Embassy. In my country, there’s almost no one who is working on these deep-tech technologies unfortunately and I am having trouble relocating due to really bad reputation of my passport which consistently ranks at the bottom of the list in terms o

Article URL: https://techoreon.github.io/verpad/canvas-playground.html Comments URL: https://news.ycombinator.com/item?id=48699646 Points: 4 # Comments: 2

Article URL: https://www.bloomberg.com/news/newsletters/2026-06-27/how-ai-powered-public-comments-could-impact-us-climate-politics Comments URL: https://news.ycombinator.com/item?id=48699641 Points: 1 # Comments: 0

<pre style='white-space:pre-wrap;width:81ex'>fix(msteams): truncate reflection prompt on UTF-16 boundary (#96578) buildReflectionPrompt truncated the thumbed-down response with String.slice(0, 500) on a UTF-16 code-unit index, so an astral character straddling the 500-char cap was cut into a lone surrogate in the reflection prompt built for the LLM. Use truncateUtf16Safe so truncation never splits a surrogate pair, keeping the existing 500-char budget and '...' suffix. Adds tests asserting the p

<pre style='white-space:pre-wrap;width:81ex'>fix(agents): keep missing tool results on current model (#95543)</pre>

Article URL: https://code.intellios.ai/cwnews/ Comments URL: https://news.ycombinator.com/item?id=48699403 Points: 1 # Comments: 0

<table> <tr><td> <a href="https://www.reddit.com/r/MachineLearning/comments/1uh61uw/hiding_messages_in_the_least_significant_mantissa/"> <img src="https://external-preview.redd.it/xL20TWLoDXtutsGuMHS1qdEyNEn6zkliHGGNaYV1H4A.png?width=640&crop=smart&auto=webp&s=2b035893689551e28412391b858ab5c0323b052d" alt="Hiding messages in the least significant mantissa bits of fine-tuned ONNX model weights [P]" title="Hiding messages in the least significant mantissa bits of fine-tuned ONNX model weights [P]"

Article URL: https://newsletter.artofsaience.com/p/the-pms-guide-to-managing-ai-debt Comments URL: https://news.ycombinator.com/item?id=48699231 Points: 2 # Comments: 0

Hi all. As someone who has spent years working with Ansible and other automation frameworks, the recent MCP boom has me fascinated. People are creating nice, typed, LLM-readable (and thus human-readable) interfaces for their servers, over a standardized protocol that exposes both tools and resources. I wanted to see if I could create a way to run scripts against these servers directly, with no AI in the loop. Ocarina lets you inspect MCP server characteristics and write Rondos (the equivalent of

RAG connects LLMs to your data and there are three different ways to do it.

Amazon's Kindle AI features help you read beyond the lines, so long as you have the right ereader.

About half of Claude users say AI can already handle 50 percent or more of their work tasks, according to a survey of roughly 9,700 users by Anthropic. In 12 months, 26 percent expect AI to cover 60 to 90 percent of their work. Early-career workers worry the most, while the heaviest users are the most optimistic about their career prospects. The article Half of Claude users say AI can already handle half their work according to Anthropic survey appeared first on The Decoder .

Article URL: https://cefboud.com/posts/llm-d/ Comments URL: https://news.ycombinator.com/item?id=48699083 Points: 3 # Comments: 0

<p> "One AI agent across your browser, tools, and messages " </p> <p> <a href="https://www.producthunt.com/products/lyto?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1182460?app_id=339">Link</a> </p>

Article URL: https://ai-creed.dev/projects/ai-whisper/ Comments URL: https://news.ycombinator.com/item?id=48698924 Points: 2 # Comments: 0

The goal was to create a highly efficient, small-scale model that can perform reasoning tasks while remaining lightweight enough to run easily on consumer hardware. Technical Stack: Base: Qwen3:4B Training: Fine-tuned using Unsloth for memory efficiency, which allowed me to run the process smoothly on an RTX 5070. Stack: Built with cu128, PyTorch, and Hugging Face Transformers. Dataset: Trained on Raymond-dev-546730/Open-CoT-Reasoning-Mini to improve Chain-of-Thought (CoT) capabilities. Comments

Ask Channel AI ( http://askchannel.ai ) allows you to quickly find relevant videos, quotes, and timestamps from any YouTube channel. Channels that are not yet indexed or imported can be done so by signing up, and imports are typically pretty fast for small to medium sized channels (~1 minute). Channels are the first-class citizen instead of individual videos, since summarization of individual videos is already a thing in YouTube. --- I spend a lot of time watching and listening to long-form cont

Article URL: https://syntheticauth.ai/posts/ai-the-falsity-of-comparison Comments URL: https://news.ycombinator.com/item?id=48698885 Points: 2 # Comments: 0

<p><em>AI agents go beyond chatbots: they use a language model to plan, call tools, and complete multi-step tasks without a human directing every action. Here is what that means for UAE enterprises, and how to build one that is production-ready and regulatory-compliant.</em></p> <blockquote> <p>UAE enterprises build AI agents by defining a precise task boundary, selecting an LLM backbone and a minimal tool surface, implementing RAG-based memory, adding an orchestration layer, and placing human-i

<p>Stop Manual Code Reviews: Build Your AI-Powered Review Workflow</p> <p>You're staring at a 2000-line PR at 4:59 PM on a Friday. Classic. Someone's asking for review feedback, and you've got about 47 other things due tonight.</p> <p>Here's the thing though: you don't actually need to read all 2000 lines yourself anymore. I've been automating my code review workflow for the past few months, and it's genuinely saved me hours every week. Not the "AI wrote all my code" hours—real, practical time s

<p>As tech firms make huge profits and investors fear losing out, both are doing their best to hold off the day of reckoning</p><ul><li><p><a href="https://www.theguardian.com/technology/2026/jun/26/openai-ai-model-release-trump-us-sam-altman-gpt-anthropic-mythos">OpenAI staggers AI model release after White House request</a></p></li></ul><p>Every couple of decades, investors will ask themselves how long can the stock market keep climbing. Is it safe to buy more shares? Is their pension or equit

A team cut their AI inference bill by more than half. Three months later, customer satisfaction was dropping and the cost savings were tied to the quality loss. Cost-optimization routing layers are a Pareto trap, and here's the detection methodology that catches them in days instead of months. The post We Built a Routing Layer to Cut Our AI Costs. It Broke the Product. appeared first on Towards Data Science .

<p>On June 18, 2026, I filed <a href="https://github.com/postcss/postcss/issues/2096" rel="noopener noreferrer">postcss/postcss#2096</a> about OIDC provenance for PostCSS. The <code>ai</code> npm account — one person, Andrey Sitnik — publishes PostCSS, nanoid, Autoprefixer, browserslist, and caniuse-lite. Combined: over 900 million weekly downloads through a single publish credential.</p> <p>Andrey's first reply was not agreement. It was a correction.</p> <h2> "CI-as-publisher increased the atta

<p><strong>AI code review has a problem: abstract roles produce generic feedback.</strong> "Saboteur" says "add error handling." "New Hire" says "this is confusing." Useful? Sometimes. Specific? Rarely.</p> <p>I built something different: a review system that uses <strong>real engineers with searchable philosophies</strong> instead of abstract roles. Linus Torvalds doesn't say "consider error handling" — he says "eliminate the special case entirely." That's not a wording difference. That's a com