All articles tagged with #llm
Firms selling AI services struggle to price them because token usage is non‑deterministic and varies with prompts, models and multi‑agent workflows; even as per‑token costs fall, total token consumption is rising, pushing vendors toward flat fees, bundles, or usage‑based pricing while awaiting howbig platform pricing shifts will affect budgets.
Google’s Android Bench adds eight new LLMs (including Claude Fable 5, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.7 Code, MiniMax M3, Qwen 3.7 Plus, Qwen 3.7 Max) and switches to the Harbor testing framework, placing Gemini 3.1 Pro in fifth behind GPT-5.4 and Claude models; Fable 5 posts the top accuracy (~84.5%) but at a high cost per run, while Gemini 3.1 Pro is cheaper (~$87) and Gemini 3.5 Flash runs the most expensive and longest; Google invites developers to run benchmarks and submit results via GitHub to shape future Android Bench tasks.
OpenAI and Broadcom introduced Jalapeño, a purpose-built ASIC designed from scratch for large-language-model inference in data centers, with early testing claiming substantially better performance per watt; development took nine months and is part of a broader effort to own more of the AI stack and reduce reliance on Nvidia, with deployments planned by year-end as the silicon race heats up.
IBM and Multiverse Computing demonstrated a quantum-classical hybrid approach that inserts Cayley-parameterized unitary adapters into a frozen Llama 3.1 8B model and runs the training-then-inference cycle on an IBM 156-qubit Quantum System Two. The result is a 1.4% reduction in perplexity with only about 6,000 extra parameters (a tiny, 0.000075% increase), enabling the hybrid model to answer certain questions correctly that the base model got wrong. The work shows a path toward quantum-enhanced AI on real hardware, but faces challenges from quantum noise and the need for future work to encode the entire quantum circuit to push perplexity further with even smaller classical needs.
ArXiv will ban authors for a year if there is incontrovertible evidence that LLM-generated content in a submission hasn’t been checked (e.g., hallucinated references or misleading meta-comments), with future submissions requiring acceptance at a reputable peer‑reviewed venue. This strengthens efforts to curb AI-generated “slop” and builds on prior policies restricting AI-heavy content to peer‑reviewed publications; bans are subject to moderator review and author appeals.
Greg Kroah-Hartman’s AI-powered bug-hunter, gkh_clanker_t1000, is surfacing and patching Linux kernel bugs across ALSA, HID, SMB, Nouveau, and IO_uring, with patches merging into mainline; the tool runs on a Framework Desktop powered by AMD Ryzen AI Max, illustrating the use of local LLMs for kernel debugging.
A peer‑reviewed study from West Coast universities finds heavy reliance on large language models (LLMs) reshapes both meaning and style in human writing. In experiments on the money–happiness question, essays written with heavy AI use were neutral far more often (69% higher) and participants produced 50% fewer pronouns with fewer personal anecdotes. LLM edits also replaced more words than human edits, often changing the essays’ meaning. Researchers warn of long‑term impacts on thought, language, and institutions, and say ideal LLMs would mirror a writer’s voice rather than overwrite it.
Ars Technica reports that Kagi Translate can output translations in playful, esoteric 'languages'—like LinkedIn Speak or Gen Z slang—by tweaking the target language, turning a translation tool into a language-play toy while highlighting both the creativity and risks of large language models and the need for guardrails.
Not in a bubble: the rise of agentic AI—where a harness guides the model and verifies results—drives sustained, higher compute demand and shifts value to integrated AI providers. Thompson traces three inflection points (ChatGPT, o1 reasoning, Opus 4.5/Codex/Claude enabling agents) and shows how enterprise adoption (e.g., Microsoft's Copilot Cowork) will amplify productivity and compute demand, while Apple leans on licensing. The result is lasting demand and fewer people needed to unlock AI's impact, making the investment case for AI capex more durable than hype suggests.
MIT researchers introduce Taming the Long Tail (TLT), an adaptive speculative-decoding framework that trains a lightweight “drafter” on idle processors to predict the outputs of large reasoning LLMs, with an adaptive rollout engine selecting the best strategy for each batch. This speeds reinforcement-learning–based training by 70–210% while preserving accuracy, and the drafter can also be reused for efficient deployment. The approach aims to reduce training cost and energy for complex AI models and has been tested across multiple models and datasets.
Boeing engineers demonstrate a space-qualified edge AI by running a compact large language model on standard hardware to autonomously analyze satellite telemetry, a development that helped BA stock rise about 2%. The story also covers a Supreme Court denial to hear a Southwest pilots’ union case, while analysts still rate BA as a Strong Buy with roughly 18.8% upside based on a $278 target after a year of gains.
Kaspersky researchers say Arkanix Stealer, promoted on dark-web forums in Oct 2025, was likely an AI-assisted, short-lived information-stealer project with Python and native C++ versions, a Discord community, and a referral scheme. It could harvest browser data (including 0Auth2 tokens), cryptocurrency wallet data, and credentials from Telegram and Discord, plus local-file exfiltration and modular plugins. The premium variant added anti-sandbox/debugging, RDP credential theft, and advanced post-exploitation tools like ChromElevator to bypass protections. The operation’s unclear purpose points to rapid, low-cost AI-driven malware development rather than a sustained campaign, with IoCs published by Kaspersky.
Mark Cuban says there are two types of AI users—those who use AI to learn everything and those who use it to avoid learning—urging companies to embrace AI while acknowledging its flaws; other tech leaders warn that overreliance on AI can deskill workers, though AI can act as a potent accelerant for career differentiation.
GCC developers are debating whether to accept patches generated by AI/LLMs like GPT-5-CodeX, with current discussions leaning towards not allowing large LLM-generated patches due to copyright and policy concerns, and a possible decision by the GCC Steering Committee pending.
A study by the BBC and European Broadcasting Union highlights that while many rely on large language models (LLMs) for news summaries, these AI tools often produce errors, with 20% of cases containing major issues, suggesting humans should still be trusted for accurate news reporting.