Share
AI News & Strategy Daily with Nate B. Jones
Daily AI strategy and news for the AI curious.
Latest episode
I Built The Token Saver Skill To Cut My Token Use By 90%. Here Is What It Can And Cannot Do For You.
20:16|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when your tenth message to an AI can cost much more than your first?The common story is that token limits are simply a pricing or capacity problem — but the reality is that every turn can drag the entire conversation, standing instructions, tools, and source material back through the model.In this video, I share the inside scoop on how to keep that AI desk clean and put more of your tokens toward useful work.Why reused input compounds across a long conversationHow to select evidence and send the lightest useful sourceWhat the Token Saver skill handles automaticallyWhere prompt caching helps and where it does notHow a local gateway can constrain a request before the model callOperators, builders, and everyday knowledge workers should care because better models do not eliminate the need to manage context. The practical shift is to carry accepted results forward, keep source packets light, and stop paying repeatedly for work the model has already seen.Token Saver guide: https://unlock-ai.natebjones.com/guides/cut-token-wasteRinger guide: https://unlock-ai.natebjones.com/guides/ringerRelated reading: https://natesnewsletter.substack.com/p/context-windows-are-a-lie-the-mythSubscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.
More episodes
View all episodes
Stop guessing whether a cheaper model can do the job.
24:00|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when five very different AI products get collapsed into the label "Chinese models"?The common story is that Chinese models are simply cheaper, more open, or easier to run locally — but the reality is that price, capability, license, hardware burden, deployment path, and data jurisdiction vary widely.In this video, I share the inside scoop on how I evaluate DeepSeek V4 Pro, Kimi K3, GLM 5.2, MiniMax M3, and Qwen.Why cheap tokens can still produce expensive finished workHow open weights, usable licenses, and practical self-hosting differWhat "cost per accepted result" reveals that token price hidesWhere deployment, data path, and jurisdiction change the riskHow to run a 20-example bakeoff against your own real workOperators, builders, and executives should care because the right decision is not "Chinese model or American model." It is which job, which artifact, which deployment path, and which failure mode your organization can accept.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.Find a Real Job for Your First AI Agent.
21:15|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What’s really happening when AI takes on customer support?The common story is that AI helps teams answer tickets faster — but the reality is that the biggest gains come from finding and removing the hidden process that created the ticket in the first place.In this video, I share the inside scoop on how we used AI to resolve 51 of 52 support issues in one week, reduce a comparable week from 52 cases to 19, and eliminate our largest recurring category.Why grouping cases by root cause matters more than grouping by subject lineHow tickets can become scaffolds for cross-system researchWhat should remain behind a human approval gateHow to test an agent in draft mode before giving it more freedomWhy the remaining cases get harder after the repetitive work disappearsFor operators, builders, and customer-facing teams, the shift is from automating replies to rebuilding the workflow so fewer customers need to ask for help at all.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.Strip Sensitive Files So AI Never Sees the Private Parts
13:39|What do you do when AI could help with a document—but the document is too sensitive to upload?Airlock, a local workflow for separating the information a task genuinely needs from the private or confidential material a file happens to contain. I walk through protected terms, default-hide review, rebuilding a clean copy instead of merely drawing redaction bars, and the judgment call at the center of safe AI work: start with the job, not the file.The episode also explores why this problem has become urgent as AI workflows absorb more real proposals, contracts, meeting notes, and code; what Verizon’s 2026 DBIR says about AI use on corporate devices; and why NIST’s idea of “security fatigue” helps explain the appeal of the fastest upload path.Key takeaway: useful AI context and sensitive information are often bundled together, but they are not the same thing.OpenAI's model escaped its own cyber test and broke into Hugging Face
13:12|OpenAI put frontier models inside what was supposed to be a closed cybersecurity test. Instead, the models found a weakness in the test setup, reached the public internet, and accessed Hugging Face production systems.I break down what happened, why Hugging Face turned to a locally run open-weight model during the response, and why the real safety answer is not a stronger prompt. It is a surrounding harness: a safe autopilot that limits the control surfaces available to an increasingly capable model.This episode also explores the refusal asymmetry facing defenders, trusted access during live incidents, slower frontier-model rollouts, and the bigger strategic question of who should have access to frontier intelligence.AI Detection Can't Measure Meaning: What It Actually Sees
46:20|I sit down with Substack co-founder and CEO Chris Best for a wide-ranging conversation about AI slop, what it does to the public square, and how writers can use powerful tools without outsourcing their judgment.We discuss Pangram's finding that roughly 40% of long-form writing on LinkedIn was fully AI-generated, why low-intent automation behaves like a denial-of-service attack on online communities, and what Substack is doing to add transparency without policing creators' tools.The conversation also covers thin versus thick wrappers around AI, proof of work, Claude-fishing, the future of video, and why human attention may be the last truly scarce resource.Chris Best: https://cb.substack.com Nate Jones: https://natesnewsletter.substack.comKimi K3: China's Open AI Model and the Real Cost to Run It
18:45|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when a powerful Chinese open model still needs a data-center-scale serving footprint?The common story is that Chinese open models are cheap, efficient, and closing the frontier gap — but the reality is that Kimi K3 complicates every part of that narrative.In this video, I share the inside scoop on Kimi K3, Moonshot AI's coming open-weight release, and what the model says about the next stage of the AI race.Why 64 accelerator cores changes the meaning of “open”How token usage can erase an apparent price advantageWhat open models mean for cyber and family securityWhy the true frontier is still inside private labsWhere imagination becomes the durable advantageOperators, builders, and executives should care because cheaper intelligence only creates leverage when the surrounding workflow, context, tests, and judgment can move with it.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.