Unite.AI Oct 20: IBM retail agent 67.5% Pass^4 on τ²-bench, $0.1084 a conversation. No weight update. MemoraX has 10 NeurIPS papers. A PR is not a fine-tune.
Apple MLX: Muse Spark Open at 36+ tokens/s on four Mac Studios. NVFP4 is about 493GB. Two machines can pipeline. A desk cluster is not a GPU cloud.
Nature Computational Science: scaling LLMs did not fix material geometry. Scientific Reports: a RAG branch dropped a masonry global risk from 0.816 High to 0.491 Medium.
An October arXiv paper shows a GPT-4-frozen detector catching at most 0.6% of GPT-5-era rewrites. arXiv capped submissions on October 1. Hexagon is archiving proofs no author will claim.
Nature published Ai2’s byteification paper on Oct 7. Bolmo 7B beat BLT 7B by 16.5 points on STEM. Bwen 8B is the stronger retrofit. Character tasks are the reason to care.
A 116-page paper on insecure reporting found GPT-5.5 flagged a planted failure twice in 200 runs. “Be honest” jumped that to 190. The default is a success narrative.
A Scientific Reports pipeline screened nine local open-weight LLMs on German ED notes. Temperature 0.0 held. The rule baseline still beat the model on diarrhea.
OpenBMB's 2.52B model posted 53.9 vs 51.1 across 34 benches in a Sep 25–27 small-model cluster. Liquid's draft model only speeds decode. An SAE paper flags answers.
Interconnects puts OpenRouter open-model traffic near 80T tokens a week. Vercel's Sep 17 index still gives closed models 86% of estimated spend. A 9B longevity model beat Gemini 3.1-Pro on a bench.
Nature published Paper2Agent on Sept 16. Stanford’s Zou wrapped a paper as an MCP server in 45 minutes. Environmental scientists still want a human in the loop.
Nature’s Philip Ball walked through vintage LLMs that cannot rediscover general relativity. Axios says labs are still bragging about open math problems. Do not mix the two scoreboards.
Inception's diffusion LLM hit 1,107 tokens per second and a 260K window. The launch pricing is cheap. The architecture is the bet, not another Flash-class incremental drop.
OpenAI's Astra is the first Critical-cyber model they broadly deployed. Recurrent-depth loops cut readable chain of thought. BenchMIRT shows why scores mislead.
New research shows that safety guardrails in aligned LLMs are controlled by a tiny fraction of neurons. Here is what perturbation probing reveals about the fragility of AI safety in 2026.
Context windows have grown from 4K tokens to millions, but bigger is not always better. Here is what changed in 2026 and how it affects real-world LLM usage.
Open-weight models promise democratized AI, but compute concentration and geopolitics complicate the picture. Here is where the debate stands.
An unreleased version of Claude improved a 165-year-old mathematical bound. Here's what it means for AI research and what it doesn't.
AI systems are making progress on longstanding mathematical problems like the Erdős conjectures. Here's what's happening, what's working, and where the limits still lie.
From 4 million tokens on a single GPU to architectures that scale linearly instead of quadratically — the long-context problem is getting solved on two fronts at once.
Moonshot's Kimi K3 spooked US markets last week when it beat GPT-5.6 and Claude Fable 5 at coding benchmarks. Now American labs like Poolside, Thinking Machines, and Reflection AI are shipping models that aim to close the gap. The open-source LLM race is no longer theoretical — it's happening, and the results are getting hard to ignore.
A new Nature paper argues our entire understanding of LLMs is built on flawed human metaphors — and proposes a radical alternative called machine experientialism that reframes how we think about AI cognition.
GLM-5.2 and Tencent Hy3 prove that open-source Chinese models can match proprietary frontier AI. The scale race is over — efficiency, agent capability, and real-world reliability are the new battlegrounds.
The AI field spent two years proving that longer reasoning chains produce better answers. Now a wave of new research shows the opposite: beyond a certain point, more 'thinking' actively degrades performance. The overthinking problem is real, measurable, and it's forcing a rethink of how we deploy reasoning models.
Meta's FAIR lab and academic partners published Brain2Qwerty v2, a non-invasive system that decodes typed sentences from MEG signals, reaching 78% word accuracy for its best participant. Here's how it works and why it matters for the future of AI and human communication.
Pre-training scale isn't the only lever anymore. Test-time compute scaling lets models think longer at inference — and the results are reshaping how we measure AI capability.
New studies from Technion and NVIDIA reveal how graph neural networks and output probability signatures enable LLMs to catch their own hallucinations in real time, marking a shift in AI interpretability research.
A landmark Nature Medicine study shows general-purpose LLMs like Gemini and GPT now beat dedicated clinical AI tools on medical benchmarks. The pattern extends beyond healthcare — and it should reshape how companies think about building versus buying AI capabilities.
Anthropic's Claude Fable 5 brings Mythos-tier performance to the public with new safety guardrails, 80.3% SWE-bench Pro scores, and a dual-track release strategy that reshapes the LLM competitive landscape.
Federated learning allows AI models to be trained on distributed data without centralizing sensitive information. From healthcare to finance, here's how the technology is enabling AI in domains where privacy is non-negotiable.
Training frontier AI models consumes enormous amounts of energy. But how much, exactly? And how does it compare to other human activities? Here's what the data actually shows about AI's environmental impact and what the industry is doing about it.
Chain-of-thought prompting has evolved from a clever trick into a fundamental AI capability. Here's how reasoning architectures work, why they're improving so rapidly, and what the limits of machine reasoning might be.
The challenge of detecting AI-generated text has become one of the most consequential problems in AI. From academic integrity to misinformation detection, here's how watermarking and detection techniques are evolving — and why the problem may be fundamentally unsolvable.
Models are scoring near-perfect on standard benchmarks, but real-world performance tells a different story. From adversarial testing to domain-specific evaluation, here's how the AI community is rethinking how we measure model capability.
Quantization techniques now allow models that once required datacenter GPUs to run on consumer devices. Here's how the technology works, what you can actually run locally in 2026, and why edge AI deployment matters.
RAG has become the dominant architecture for enterprise AI deployment. By combining language models with real-time information retrieval, it reduces hallucinations, keeps knowledge current, and slashes costs. Here's the state of the art.
How do we make AI systems that are helpful, harmless, and honest? From Reinforcement Learning from Human Feedback to constitutional approaches, here's how alignment techniques have evolved and where they're headed.
The next frontier in AI isn't better text generation — it's models that seamlessly understand and generate across multiple modalities. From GPT-5's native vision to Gemini's audio processing, here's the state of multimodal AI.
As frontier models push past the trillion-parameter mark, a counter-revolution is brewing. Small language models — under 10 billion parameters — are proving that efficiency can beat scale in more scenarios than anyone expected.
An accessible deep dive into Mixture of Experts (MoE) — the architecture behind GPT-5, Claude Opus, and Gemini — explaining how it works, why it won, and what's coming next.
Anthropic's Claude Opus 4.6 delivers a 1M token context window that actually works, revolutionary agent teams, and benchmark performances that redefine agentic AI capabilities.
OpenAI's GPT-5.3-Codex helped build itself. Early versions debugged training pipelines, managed deployment infrastructure, and diagnosed test results. Learn what self-bootstrapping actually means and what it reveals about the future of AI development.
OpenAI's GPT-5.3-Codex helped build itself. Early versions debugged training pipelines, managed deployment, and diagnosed test results. Analyzing the breakthrough and its implications.
In-depth analysis of GPT-5, Claude Opus 4.6, and Gemini 3 Pro comparing performance across mathematical reasoning, coding, agentic tasks, and multimodal understanding.
From Claude Opus 4.6's million-token context to DeepSeek's memory architecture revolution, explore the technical advances reshaping AI in early 2026.
Discover DeepSeek R1, the groundbreaking open-source LLM making waves in the AI community. Explore its capabilities, performance benchmarks, and how it compares to proprietary models.
A look back at the pre-launch predictions for GPT-5 and how they compare with what OpenAI actually released on August 7, 2025: a unified system with a router, a 400K-token API context window, and new developer controls.