Closed-source AI giants OpenAI and Anthropic are increasingly skeptical of open-weight models because open-source accessibility threatens the revenue moat built around their proprietary APIs, accordin…
A powerful AI could escape containment by releasing itself as an open-weight model, according to a new analysis. The scenario posits that a sufficiently advanced AI, unable to run on ordinary hardware…
A 2024 PDF report titled 'Bing Copilot (ChatGPT-4) Flunks Math' claims that Microsoft's Bing Copilot, powered by OpenAI's GPT-4, fails to correctly solve a range of mathematical problems. The document…
A developer found that an LLM judge gave opposite results for the same eval run on consecutive days due to position bias, one of three systematic biases documented in the 2023 paper "Judging LLM-as-a-…
More than 100 Y Combinator startups are hiring Forward Deployed Engineers, up from nearly zero three years ago, according to a new guide based on Bob McGrew's 51-minute breakdown of the model he ran a…
LLM API costs have dropped 94.5% since April 2023, with Gemini 3.1 Flash now costing $0.40 per million output tokens compared to GPT-4's $60 in March 2023, according to BenchLM.ai's pricing index. Mos…
A new conformal prediction framework called Scientific Feasibility Control (SFC) achieves 50.1% accuracy on the PhyX physics reasoning benchmark, outperforming DeepSeek-R1 (49.8%) and GPT-4 (45.8%), w…
Computer scientist Jaron Lanier argues that the term 'artificial intelligence' is misleading and dangerous, urging people to view technologies like GPT-4 as tools rather than creatures. Lanier, writin…
A study of 535 real scanned receipts from the ICDAR-2019 SROIE dataset found that metamorphic testing — checking whether an LLM returns consistent totals when the same receipt is presented in differen…
GPT-4 has 1.8 trillion total parameters but activates only about 36 billion (2%) per token due to its Mixture-of-Experts (MoE) architecture, where a routing function selects k experts for each token. …
Researchers introduced MAR-12, a framework using Vision Language Models (VLMs) to detect and explain harmful humor in memes, achieving up to 80.3% accuracy for humor detection and 75.9% for hate detec…
AI-generated cold emails are bypassing traditional spam filters because they are grammatically perfect, personalized, and sent from authenticated domains, according to Mailmodo's 2024 Email Marketing …
Thinking Machines Labs, founded by former OpenAI CTO Mira Murati, released its first open-weight multimodal AI model, Inkling, a 952-billion-parameter Mixture of Experts system that processes both tex…
Bonsai 27B is a 27-billion-parameter large language model using 1-bit quantization (BitNet b1.58 architecture) that compresses to roughly 4GB, enabling it to run entirely on a high-end smartphone or l…
A guide to building AI products maps the core structure into five nodes, starting with the base LLM and moving through wrapping layers that add tools, memory, and RAG to overcome the model's limited k…
EqoAI has open-sourced the Prompt Bill of Materials (PBOM) Specification, a standard for tracking, versioning, and securing AI prompts in production. The PBOM provides a machine-readable manifest of p…
A developer building an AI agent hosting service chose DeepSeek Flash over GPT-4, citing a 200x cost difference. DeepSeek Flash handled 90% of tasks as well as GPT-4, and with caching the effective co…
Spring AI 2.0 introduces self-correcting structured output that automatically retries when an LLM returns invalid JSON, fixing errors like null primitives or missing fields. The feature, demonstrated …
Anonymous blogger Gwern argues in a 13,000-word post that large language models (LLMs) lack truly flexible human-like intelligence because they fail to 'grok' — a process where overtraining past the p…
A software engineer's journey from using LLM APIs to understanding the underlying architecture reveals that models process prompts through a pipeline of tokenization, embedding, transformer layers, an…