ResearchAudio AdminAdmin•Thursday•NewslettersA 56-Layer Network Lost to a 20-Layer One on Its Own Training DataResNet’s answer was a plus sign that teaches every layer how to do nothing. A decade later, every LLM you call still runs on it.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Tuesday•NewslettersGoogle's object-oriented harnessWithout the First Run, Google's Harness Router Drops 12.64 Pointsvia ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Oct 22•NewslettersDeepSeek released its AI engine but kept its storageThat storage lets it reuse data for 95 of every 100 tokens and keep its costs lowvia ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Oct 19•NewslettersSame Weights, New Files: GLM-5 Scored 11.5 Points HigherChanges with no test run decreased a code benchmark result from 73.4 to 63.2.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Oct 14•NewslettersCodex Found the Right Number, Then Used the Wrong OneMemory alone did not fix this. Prompt rules that steer the search did.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Oct 13•NewslettersGPT-5.4 Obeyed “No Reasoning” and Lost 28.5 PointsThinking off still reasons in 99.69 percent of open answers. The fix is the question.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Oct 12•NewslettersYour LLM server was built for one model. Token routing needs two.A NeurIPS 2026 paper from Tsinghua's thu-nics group explains why token-level routing has been stuck in notebooks, and ships a serving system that makes it 2–64× faster than what researchers had been running.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Oct 6•NewslettersCan We Still Watch AI Think?Chain-of-thought provides insight into AI reasoning. New research shows where that window starts to break.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 29•NewslettersSonnet 5.5 kept the price. Your bill could shrink.Faster output, fewer tokens, and the API changes to check before switching.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 24•NewslettersThe cheaper model can cost more Opus 5.5 and GPT-6 Sol, explained through five diagrams about agents, caching, and the cost of finished work. via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 22•NewslettersGrok 4.7 Beats a Grok 4.6 That Was Corrected in AugustThree regressions shrank by revision. Opus 5 left the coding charts. Biology fell by design.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 18•NewslettersYour AI agent says it’s done. Is it?Designing the system that assigns the work, checks the result, and knows when to stop.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 17•NewslettersGoogle's OKF Needs One Field. Version 0.2 Locks the SQL.Most explainers still describe v0.1. We fed the reference attester five fake receipts.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 16•NewslettersReviewers Caught ~30 Dropped Names. A Script Found 1,703.One million lines refactored by 1,393 agents. The receipts Nous published second.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 14•NewslettersThree Rivals Agreed. Two of the Three Steps Need Washington.Anthropic, OpenAI and xAI landed on the same position in one afternoon. The next morning, the people who would have to write it into law said no.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 13•NewslettersThree CVEs for One File. The Fourth Has No Name.Eight findings across seven agents. The command runs before any model call.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 12•NewslettersAnthropic Says Slow Down. The One Promise Is a Desk and a Badge.Dario Amodei's September essay asks every frontier lab to pace how fast models improve. Read it line by line and one of its three steps is a commitment, two are requests, and nothing in it says how much slower, or from when.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 11•NewslettersDeepSeek's New Base Loses 12 of 16 Rows to the Model It ReplacesDeepSeek claimed to beat Anthropic and OpenAIvia ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 9•NewslettersTencent's World Generator Runs on Claude Opus 4.8GPT-Image-2 paints objects onto a render, SAM3D lifts them. The evaluation has no numbers.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 7•NewslettersAstra and Fable 5.1 Share a Rate Card, Not a BillSame 10 in and 50 out per million tokens. One burns 3.3x the tokens per index task and bills 2.4x. One cache line sits 4x apart.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 5•NewslettersWhy AI agents keep relearning the same lessonsA study turns software knowledge into tested instructions for research agents. The gains are substantial. Understanding where they come from takes a closer look.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 4•NewslettersOpenAI's Most Aligned Model Also Writes the Least DownScope violations fell to 0%. Reasoning monitorability fell too. Your agent can get a 403.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Sep 2•Newsletters5 Trillion Context. Zero Tokens.Anandkumar's physics model is graded by the equation itself, not by labels.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Aug 31•NewslettersA 66-Cent Draft Caught 5 of 36 FabricationsThe full stack reached 33, at twelve times the cost and 108 times the tokens.via ResearchAudio👍❤️😂🔥💯
ResearchAudio AdminAdmin•Aug 29•NewslettersThe Draft Cost 0.11M Tokens. The Checking Cost 11.8M.Fabrication detection went from 5 of 36 to 33 of 36. What each layer bought.via ResearchAudio👍❤️😂🔥💯
