443 summarised stories about Large Language Models, each linking back to the original source. Browse all topics →
TechCrunch 3 weeks ago 36
A pair of cryptanalysts says OpenAI’s Astra and Anthropic’s Claude Opus 5 decoded multiple long-unsolved Enigma messages using LLMs. One Astra result recovered plaintext of a message that had baffled researchers since 2005. This leaves just seven unbroken Enigma messages (plus one with known plaintext), and more may be broken as models are reused against remaining records.
Deep Learning Weekly 3 weeks ago 18
Deep Learning Weekly published Issue 474 covering new deep-learning and AI agent releases, benchmark findings, and research papers. Pricing and performance updates include Anthropic’s Claude Opus 5.5 at $4/$20 per million tokens, 40% below Opus 5, with a 66.4% Terminal-Bench 4.0 score and output over 30% faster. The issue adds new references for model/tool adoption and introduces research like RRSI that aims to improve agent-harness self-improvement while limiting overfitting and reducing policy-token use by 30%.
MarkTechPost 3 weeks ago 6
Perplexity Research trained a model in its Computer agent using real user sessions that included failures, applying rejection-sampling fine-tuning plus hint-guided self-distillation. In a live A/B test, tool-call failures fell from 2.24% to 1.77% across two trained checkpoints, a reported 21.2% relative reduction. The improved checkpoint runs only as an internal option inside Perplexity Computer and the post-trained weights or training code were not released, so it is not directly deployable elsewhere.
StackSweep 3 weeks ago 15
TypeSafe JEV 1.13, a hosted decision-only LLM judge, matched GPT-6 on routine preference and evidence-grounded factuality but not on hard correctness tests. It cost $0.044 per 1,000 judgments versus GPT-6 at $12.18 (about 277x cheaper). Using JEV’s calibrated confidence to run a cascade cut spending by escalating 34% of items to GPT-6 while reaching 91.3% accuracy, but it still lagged on JudgeBench hard correctness by 14.6 points.
Astral Codex Ten 3 weeks ago 5
AI 2027 debates raised concerns that a “neuralese recurrence” design could let models think harmful ideas without leaving detectable English chain-of-thought. The article points to Jakub Pachocki’s claim that Astra’s computation depth is within a factor of two of GPT-4. The discussion shifts from fear of unmonitored looping to a more technical argument about how recurrence changes simulated vs real layer depth and why chain-of-thought bottlenecks may still limit alignment risk, while acknowledging “true neuralese” is not yet realized.
The daily briefing
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.
TLDRocket
18 October 2026
AI oversight stopped being a PowerPoint exercise today and got hands-on. Anthropic said it has given outside AI watchdog METR “employee-like” access, embedding a METR team in its offices with access to hardware and internal systems so risk evaluation can run closer to the actual model and infrastructure—not just post-hoc reports. That trend of tightening control shows up elsewhere: U.S. banks and brokerages are blocking outside AI agents from logging in and moving money, with Charles Schwab reporting it blocks agents “in the tens and twenties each week,” while routing limited balance viewing and analysis through Schwab’s own MCP servers. Even the courts are joining in—at least eight states sued OpenAI, alleging ChatGPT helps “cyberhacking agents” and misleads adults, raising the stakes for how tightly companies must constrain agent behavior.
Read the full briefing →