Claude Opus 4.8 enhances collaboration with improved judgment and efficiency, featuring dynamic workflows for large-scale tasks and a fast mode that is now three times cheaper than previous models.
The Dead Economy Theory
The Dead Economy Theory posits that AI-driven automation threatens to eliminate human labor, leading to a market where the very consumers needed for economic stability are rendered obsolete. This shift is driven by the need for AI companies to justify their massive valuations, which rely on replacing human workers with machines.
Anthropic raises $65B in Series H funding at $965B post-money valuation
Anthropic has secured $65 billion in Series H funding, elevating its valuation to $965 billion, with significant backing from major investors like Altimeter Capital and Sequoia Capital, aimed at enhancing Claude's capabilities and scaling operations. Link to article
Claude Code – Everything You Can Configure That the Docs Don't Tell You
Claude Code's source code reveals undocumented features such as a YOLO Classifier for auto-mode permissions, allowing users to configure safety decisions with plain English descriptions, enhancing the AI's autonomy and flexibility.
Liquid AI reveals 8B-A1B MoE trained on 38T
LFM2.5-8B-A1B enhances on-device AI with a 128K context window and 38T tokens in pretraining, significantly improving performance in reasoning and task execution on consumer hardware.
Real-time LLM Inference on Standard GPUs: 3k tokens/s per request
Kog AI's Inference Engine (KIE) achieves 3,000 tokens/s on 8× AMD MI300X GPUs and 2,100 tokens/s on 8× NVIDIA H200, demonstrating that standard datacenter GPUs can match dedicated inference hardware speeds through optimized software stacks.
Orchestrating AI code review at scale
AI code review at Cloudflare utilizes a composable plugin architecture with up to seven specialized agents to enhance accuracy and efficiency, significantly reducing median review time to 3 minutes and 39 seconds across 48,095 merge requests in the first month of operation.
A new dataset with more that 100M hi-quality, curated images, with captions and meta data!
MONET is a new, open dataset containing 104.9 million high-quality images curated from a pool of 2.9 billion, complete with captions and metadata, and is licensed under Apache 2.0.
Even (very) noisy LLM evaluators are useful for improving AI agents
Noisy LLM evaluators can still effectively rank AI agents, as their average scores can indicate which agent performs better overall, despite individual output inaccuracies. This insight reveals that even evaluators with low output-level correlation can be useful for offline selection processes, allowing for the deployment of improved AI agents over time.
ATLAS: Autoformalized Textbook Library At Scale
ATLAS is a Lean 4 library that autoformalizes textbook mathematics using LLMs, encompassing diverse fields such as algebra, geometry, and theoretical computer science to create reusable formal building blocks for future formalization efforts.
Making LLMs tell you how confident they really are through probe-targeted fine tuning.
Probe-targeted fine-tuning (LoRa) enables LLMs to express their internal confidence levels, achieving an AUROC of 0.76–0.88 in distinguishing correct from incorrect answers, despite often reporting 99% confidence inaccurately.
CVE-Bench: testing LLM agents on real-world vulnerability patches
No AI model consistently fixes security vulnerabilities, with the best performer achieving a 50% overall solve rate and 60% under optimal conditions, highlighting significant limitations in current AI capabilities for real-world security tasks.
When Should Models Change Their Minds? Contextual Belief Management in Large Language Models
Contextual Belief Management (CBM) is essential for language models to effectively manage information over long interactions, balancing updates and preservation of state while filtering out noise.
Wall-OSS-0.5: 4B VLA with open training code and zero-shot real-robot evaluation
Wall-OSS-0.5, a 4B VLA from X Square Robot, showcases a novel evaluation method by testing pretrained checkpoints on real robots prior to fine-tuning, achieving impressive results with zero-shot performance on a 17-task suite, including a notable 82% on Rope Tightening.
Building a monokernel for LLM inference on AMD MI300X - up to 3,300 output tokens/s per request
The monokernel developed for LLM inference on AMD MI300X achieves an impressive 3,300 output tokens/s per request, leveraging optimizations that align memory access with the die topology.