The Archive

User achieves 50 tokens/sec with Qwen 3.6 27B on RTX 3090 using MTP speculative decoding at 100k context.

Reddit post urging opposition to GUARD Act, which would mandate ID/biometric verification for all AI chatbot access in the US.

Reddit post claiming Claude made false medical credentials claims; anecdotal observation without verification or systemic analysis.

Reddit user criticizes Anthropic for perceived inconsistency between military use ethics policies and data handling via third-party infrastructure.

For me and my business, 4.6 was the bee's knees. We fired OPEN AI, stopped using GPT in process tasks and moved a lot of our automation and workflow into 4.6. Today we went back to 4.6. 4.7 is burning us out in checks and balance. Its **WAY TO AGGRESSIVE** in making it's own decisions, moving forward with bad direction. What we missed was "before I continue" and some checks and balances. We burn context, tokens, credit, and tool usage insanely fast with 4.7 with about 50% error rate. Has anyone experienced this? I just did a switch to 4.6 3 hours into a large task that kept failing with 4.7...

Deal follows others with Microsoft, Amazon, and more.

Barry Diller defended OpenAI CEO Sam Altman, while warning that AGI remains an unpredictable force needing guardrails.

MTP speculative decoding ported to Qwen 3.6 35B shows modest 2.5-6% speedup vs. 2-2.5x on 27B; architecture may limit gains.

TSMC backs renewables during record demand for energy-hungry chip manufacturing.

The deal, announced last November, would have seen Perplexity's AI search engine integrated directly into Snapchat.

xAI's real business may be more about building data centers than training AI models.

Google has pulled the plug on Project Mariner, an experimental feature designed to perform tasks for you across the web, as reported earlier by Wired's Maxwell Zeff. The Project Mariner landing page now contains a message that says: "Thank you for using Project Mariner. It was shut down on May 4th, 2026 and its technology voyaged to other Google products." Google first revealed Project Mariner in December 2024 and later announced an update allowing it to perform up to 10 tasks at a time. Over the past year, Google has integrated features powered by Project Mariner into its other AI tools, inc...

Trump forced to admit Biden was right on AI safety testing.

Reddit speculation that xAI will dissolve as a separate entity; unconfirmed claim lacking official sources or detail.

Anthropic product lead outlines three near-term model focus areas: improved autonomous coding capability, extended context windows with memory, and multi-agent coordination.

User reports successful local inference with Qwen3.6-35B on AMD R9700 GPU, generating functional code and tests via llama-cpp.

GB10 Solution Atlas, Rust+CUDA inference engine, achieves 100+ tok/s on Qwen 35B; open source, minimal footprint, no Python runtime.

Hi, A few hours ago we started seeing unauthorized charges coming out of our company card. This card was only provided when purchasing Claude Max. By sheer luck, we blocked the card after the first transaction. Since then there have been 5 more attempted charges. All of them were to random services like Auto Glass and Walmart, most showing Memphis in the header. Check your cards.

AI and Crypto Czar David O. Sacks speaks during a meeting of the White House Task Force on Artificial Intelligence Education at the White House. | Matt McClain/The Washington Post via Getty Images Hello and welcome to Regulator, a newsletter exclusively for Verge subscribers about tech, politics, and Washington intrigue. (It's basically House of Cards, but for nerds.) Not a subscriber yet? You really should become one, and to save you a Google search, here is the direct link to do so! And do you think I should know something? Send it to tina.nguyen+tips@theverge.com. On Monday, The New York T...

Reddit user reports Claude exhibiting erratic behavior; anecdotal observation without technical detail or reproduction steps.

Reddit discussion argues prefill latency is underemphasized vs. token generation speed in local LLM benchmarking and optimization focus.

User complaint about Claude Opus 4.6 performance restrictions and usage limits; expresses concern about competitive loss to alternative tools.

Zyphra releases ZAYA1-8B, an 8B parameter model optimized for inference efficiency, trained on AMD hardware.

Since two days ago quality of responses dropped dramatically, and now I see why. This is Sonnet 4.6. At this point it's not even trying.

Claude declined to optimize a CV for Philip Morris tobacco role, citing ethical concerns about tobacco marketing.

Just a reminder that the data centre announced to be used is the one xAI installed a massive amount of toxic gas turbines to power it, which is illegal and deadly to the local area.

Genesis AI claims Gene'26.5 is autonomous; limited details available from social media post.

Reddit post speculating on Anthropic-SpaceX partnership; lacks concrete details or sourced reporting.