Share
AI News & Strategy Daily with Nate B. Jones
Your Prompts Didn't Change. Opus 4.7 Did.
•
What's really happening inside Claude Opus 4.7 when Anthropic ships their smartest model ever into a week where OpenAI pushed the biggest Codex update since launch and everyone is racing toward IPO?
The common story is that 4.7 fixes the quitting problem from 4.6, but the reality is that this is a directed optimization with a new tokenizer that maps the same prompts to up to 35% more tokens, and the model went backward on web research while surging on enterprise knowledge work.
In this episode, I share the inside scoop on whether 4.7 is worth the upgrade:
• Why the persistence fix is real but comes with a combative literalism that punishes vague prompts
• How a 465-file adversarial migration test exposed trust failures in both frontier models
• What Claude Design reveals about Anthropic competing on harnesses, not just models
• Where the economics are heading when serious work gets serious tokens and casual interactions do not
Leaders who migrate without benchmarking their specific workflows will discover that Browse Comp dropped from 83 to 79 and terminal execution trails ChatGPT 5.4 by nearly 6 points.
Subscribe for daily AI strategy and news.
For playbooks and analysis: https://natesnewsletter.substack.com/p/opus-47-is-smarter-more-literal-and?r=1z4sm5&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
More episodes
View all episodes
Prove Your Value at Work in the AI Era: Judgment Artifacts
10:33|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when AI makes everyone's work look polished?The common story is that AI makes people more productive -- but the reality is that it also makes old evidence less trustworthy.In this episode, I share the inside scoop on how to prove you are good at work when outputs are easier to generate than ever.Why portfolios are no longer enough on their ownHow whiteboard-style conversations reveal judgmentWhat situation, decision, risk, and change show about real workWhere Talent Board-style evidence fits into careers and hiringHow to make your reasoning visible without over-performingIf you hire, manage, build, or are trying to grow into a new role, the shift matters: the scarce signal is no longer just what you produced. It is whether people can see how you understood the problem, handled tradeoffs, and improved the work.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.How I AI: My Weekly Codex Experiments
05:39|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when AI stops being a chat box and starts becoming a working context system?The common story is that better prompting is about clever wording — but the reality is that the work is moving toward cleaner context, better task shape, and agents that can stay oriented through long runs.In this video, I share the inside scoop on how I'm using AI this week: assembling context windows, using Codex on local files, and shifting from prompt engineering into collaborative task definition.Why local folders can become clean context windows How Codex changes long document, spreadsheet, and code workflows What changed in prompting after agentic workflows got better Where Claude still fits for polish, salience, and design Why multi-threaded drafting now feels practicalFor operators, builders, marketers, and executives, the important shift is not just which model wins. It's learning how to structure the work so the model can help you think, execute, review, and iterate.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.Product Management When Software Creation Is Cheap
12:37|For deeper playbooks and analysis: https://natesnewsletter.substack.com/Product management is changing as AI makes first versions cheaper. The obvious advice is that PMs should prototype more, but the deeper shift is about judgment: deciding what should exist, what should be deleted, who a product is for, what standard it needs to meet, and what the company is willing to rely on.Nate walks through the move from rationing scarce engineering to classifying software abundance, including the Prototype Commons, production class ladders, and why promotion and demotion become core product work.Hosted on Acast. See acast.com/privacy for more information.Agent Product Analytics: What Your Dashboard Can't See
11:50|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when your user is no longer just clicking, but delegating work to an agent?The common story is that agent failures are engineering incidents — but the reality is that many of them are product analytics failures hiding inside the agent run.In this episode, I share the inside scoop on why product teams need a new analytics layer for agent products.Why chat logs are not enoughHow agent runs replace sessions as the unit of behaviorWhat Salesforce's Agent Work Units signal about SaaS metricsWhere completion, acceptance, and correction rates fitWhy product analytics becomes the rudder for agent autonomyOperators, product leaders, and builders should care because agents move too fast for old dashboards. If you cannot see intent, tool calls, permissions, corrections, completion, and trust in one run-level view, you are steering with missing instruments.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.How to verify AI-generated Office files before they ship
19:28|For deeper playbooks and analysis: https://natesnewsletter.substack.com/AI can make PowerPoint decks, Excel workbooks, and Word documents faster, but faster is not the same as trustworthy. In this episode, Nate breaks down a practical workflow for AI Office files: prepare the sources, define the structure, constrain the artifact creation, and verify the output like a skeptical reviewer.The key idea: the file is not the whole thing. The file is the visible output of a knowledge-work system. If the claims, numbers, sources, assumptions, charts, and formulas cannot be traced, the artifact may look finished while quietly breaking trust.Hosted on Acast. See acast.com/privacy for more information.Public AI Work: How Teams Actually Learn From AI
16:24|For deeper playbooks and analysis: https://natesnewsletter.substack.com/What's really happening when AI work moves out of private chats and into shared company spaces?The common story is that AI adoption is mostly about buying better tools -- but the reality is that the companies learning fastest are making the work itself visible.In this episode, I share the inside scoop on how public AI workflows can become apprenticeship infrastructure for teams learning to work with agents.Why Slack is becoming a practical substrate for human-AI collaborationHow Shopify's River workflow makes agent work observableWhat most companies lose when AI work stays hidden in private windowsWhere senior operators should make non-sensitive AI work publicWhy constraints can turn AI use into shared learning instead of isolated productivityThis matters for operators, builders, executives, and team leads who need AI adoption to compound across the organization, not just live inside the habits of a few early adopters.Subscribe for daily AI strategy and news.Hosted on Acast. See acast.com/privacy for more information.AI Agents Create a Hidden Platform Team Bottleneck
46:36|What's really happening inside an AI infrastructure team when agents start doing the work? The common story is that AI makes every team faster. The reality is more complicated, because the speed arrives unevenly and someone underneath has to absorb it. I sat down with Emma, who leads data infrastructure engineering at OpenAI, to find out what her team is actually building to stay ahead of the agents.In this interview, I share the inside scoop on why platform teams become the bottleneck when AI agents scale across a company:- Why app teams and platform teams accelerate at completely different rates- How goal-directed agents start to feel adversarial without meaning to- What OpenAI's data platform team built to buy back time- Where a private eval suite fits into surviving constant model upgradesFor platform and infra engineers, this is the telegraph from the future: the pinch point is coming, and the teams that instrument the load now are the ones who stay standing.For deeper playbooks and analysis: https://natesnewsletter.substack.com/Why Big Tech Now Runs an AI Factory
23:36|What's really happening inside the AI supply chain that powers every model you use?The common story is that AI is a software business with a fancy backend. The reality is more complicated, and it changes how you should buy, budget, and contract for AI.In this podcast, I share the inside scoop on why your AI vendor contract is now a supply contract in everything but name: • Why "capacity constrained" points to memory and packaging, not GPUs • How hyperscaler CapEx reshapes every vendor agreement you sign • What questions belong in your next AI investment review • Where a single supply chain delay stops you from shipping AIFor operators and CFOs, the takeaway is sober: cheaper tokens are real and serving costs keep falling, but the industrial base underneath your AI strategy still demands supply assurance, utilization discipline, and contracts that account for allocation risk.Subscribe for daily AI strategy and news.For deeper playbooks and analysis: https://natesnewsletter.substack.com/AI Project Room: Organize Files Before Asking AI to Write
21:50|Now I have the full transcript. Building the deliverable.What's really happening when prestigious law firms file motions full of AI hallucinations?The common story is that better prompts prevent hallucinations — but the reality is more complicated.In this video, I share the inside scoop on the project room workflow that makes hallucinations structurally unlikely: • Why your first AI prompt should never be "do the thing" • How agents now walk folder trees and compare files cleanly • What artifacts make an agent's judgment visible and inspectable • Where most serious knowledge work breaks down before the draftOperators doing high-stakes knowledge work with AI agents need to shape the canvas before the writing starts, or they ship the same soft spots that landed Sullivan and Cromwell in front of a federal judge.Subscribe for daily AI strategy and news. For deeper playbooks and analysis: https://natesnewsletter.substack.com/