PHP’s disclosure dispute and the responsibility to users
Rejecting a social contract does not resolve a publisher’s responsibility to users still waiting for a usable fix.
Article archive
Find Roundup Tech articles by model, company, or technical question.
32 articles
Your browser extension wants every site: how to limit its access
Understand extension warnings, restrict website access in Chrome, Edge and Firefox, and check private browsing and sync settings.
Full Wi-Fi bars, slow mesh: find the weak link before moving a node
Separate internet speed, device Wi-Fi and mesh backhaul with a wired baseline, careful placement tests and model-specific Ethernet checks.
An AI answer cites your PDF. Check the sentence before you reuse it
Trace PDF citations to the original page, check quotes and table cells, and separate a useful source pointer from support for an answer.
Local LLM memory: Leave room for the conversation
Model weights are only part of local LLM memory use. Learn how context, KV cache, quantization, and CPU offloading affect the budget.
Before you send a file link, check who can actually open it
Check link audience, editing rights, inherited folder access, and what remains after you stop sharing.
A USB-C cable charges your laptop but the monitor stays blank: what to check
Check the computer port, cable, display input, and dock separately before buying replacement hardware.
Before deleting cloud files, check what your backup can restore
Cloud synchronization, device backup, version history, and a separate archive solve different problems. Test the recovery path before relying on it.
Changing phones with passkeys: check your recovery route before you erase the old device
A passkey can make sign-in easier, but changing devices still requires knowing where the credential lives and how to recover the account.
Measuring Mercury 3 beyond token throughput
Parallel decoding can shorten model waits. The useful benchmark is how much of that saving survives an entire task.
Claude Opus 5.5 cuts API prices as GPT-6 Sol launches
Anthropic and OpenAI lowered API prices on the same day. Cache use, reasoning effort, and retries still affect the cost of a completed task.
Grok Build and ZCode open their source after workspace upload incidents
The releases allow client-code inspection. Audits, reproducible builds, and differences from shipped editions still need scrutiny.
Reasoning costs in DeepSeek V4.1-Flash and GLM-5.3-Flash
These Flash releases prioritize reasoning over immediate replies. Internal reasoning tokens affect latency and the bill, even when API rates look low.
DeepSeek V4-Flash's lower test costs come with a performance tradeoff
DeepSeek V4-Flash's lower test cost comes with lower benchmark scores than premium models. Developers can route tasks according to cost and capability.
AI workloads on an unlimited token budget
An unlimited token budget would allow longer agent runs and larger prompts. Model access and the cost of self-hosting would still limit what can run.
GPT-5.6's reported access limits and GLM-5.2's open weights
The Information reports a staged GPT-5.6 release with government partner reviews. GLM-5.2 weights are available under an MIT license.
AI-designed experiments in chemistry and materials research
Robots learn from human video, while AI agents propose chemical experiments and screen materials. Laboratory tests remain necessary to validate the results.
US export controls restrict access to Anthropic's Fable 5 and Mythos 5
US export controls restrict Fable 5 and Mythos 5 by nationality, while Kimi K2.7 and GLM-5.2 make weights available for download.
Anthropic’s report on AI-led research and international coordination
Anthropic calls for international coordination on AI-led research while supporting US chip export controls aimed at China. The report also discusses safety-review limits.
DeepSeek makes V4-Pro's 75% discount permanent
V4-Pro's permanent API prices undercut several providers hosting the same open-weight model. Cached-input pricing also affects costs for agent workloads.
OpenAI reports an AI-generated counterexample to Erdős's unit distance conjecture
The model searched for counterexamples to the unit distance conjecture. Its argument was checked by mathematicians but has not been verified in Lean or Coq.
Gemini 3.5 Flash costs more than 2.5 Flash as Google adds Omni
Gemini 3.5 Flash improves on Google's older models at a higher price than 2.5 Flash. I/O also brings Omni Flash and a Gemini CLI retirement plan.
AI tools and hiring cuts in software, banking, and professional services
Hiring and staffing data from software, banking, consulting, law, and accounting, alongside AI tools and the debate over education and job skills.
OpenAI goes multi-cloud; SpaceX absorbs xAI and supplies Anthropic
OpenAI changes its Microsoft agreement, xAI becomes SpaceXAI, and Anthropic secures Colossus capacity.
Grok 4.3 compared with DeepSeek V4 Flash, MiniMax M2.7, and MiMo 2.5
Published benchmarks, API prices, and deployment options for Grok 4.3, DeepSeek V4 Flash, MiniMax M2.7, and MiMo 2.5.
Warp releases its terminal source under AGPL with OpenAI sponsorship
Warp's release pairs an open terminal client with a plan to attract contributors as AI coding tools increase competition.
Xiaomi demonstrates MiMo V2.5 Pro on extended coding tasks
Xiaomi's demos include a compiler and a video editor built over extended tool-use sessions, with reported recovery from failed tests during the compiler task.
DeepSeek V4's Huawei hardware support and open-source release
DeepSeek V4's MIT license allows commercial deployment, while Huawei Ascend support offers an alternative to NVIDIA. Hardware and software maturity still matter.
DeepSeek V4's architecture, pricing, and deployment options
DeepSeek V4 combines sparse activation, compressed attention, and downloadable weights. Pro and Flash offer different costs and deployment requirements.
How LightRAG combines graph indexing with entity and topic retrieval
LightRAG's published comparisons include NaiveRAG and GraphRAG. Graph construction adds indexing costs, and retrieval gains vary by dataset.
GPT-5.5 scores 82.7% on Terminal-Bench 2.0
GPT-5.5's published results show a lead on Terminal-Bench 2.0 and closer competition on coding and reasoning tests such as SWE-Bench Pro and HLE.
OpenAI’s Privacy Filter detects personal information locally
OpenAI's open-weight Privacy Filter detects personal information for redaction. Accuracy varies with the domain and the amount of context.
No matching articles. Try a shorter query or return to all articles.
All articles →Subscribe via RSS
Copy the feed address below and add it to your RSS reader.
Add this address to your RSS reader to receive new articles.