Build or Replaced
Today: NotebookLM is now Gemini Notebook | LM Studio Bionic: the AI agent for open models | Decoy Font Episode date: 2026-07-17.
Download MP3 →JOSH: It's Friday, July 17. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: The theme today is simple. Models are getting bigger, agents are getting local, and humans are becoming the bottleneck. JOSH: Stick around. Erik's got an AI pro tip at the end about making agents prove their work before you trust them. [pause] JOSH: First headline. NotebookLM is now Gemini Notebook. Is that just a rename, or does it matter? ERIK: It matters because Google is cleaning up the product story. NotebookLM was a great tool with a weird island name. Gemini Notebook says, this is part of the main AI stack now. JOSH: Second one. LM Studio Bionic, an AI agent for open models. Local agents are back? ERIK: They never left. They just stopped being toys. If LM Studio can make open models act like real agents on a desktop, that's a huge deal for teams that can't send everything to a hosted API. JOSH: Third headline. The human-in-the-loop is tired. That sounds like half the people listening. ERIK: Yeah, because most human-in-the-loop systems are just alert fatigue with better branding. If your AI escalates every weird edge case to a person, you didn't build a workflow. You built a pager. [pause] JOSH: Let's start there. The human-in-the-loop is tired. That one feels very Build or Be Replaced. ERIK: Completely. Human-in-the-loop is one of those ideas that sounds responsible, and sometimes it is. But people use it as an excuse to avoid building the hard part. [beat] ERIK: The hard part is deciding what the system can do alone, what it should block, and what actually deserves a human. JOSH: Isn't the human supposed to be the guardrail? ERIK: Sometimes. But if the human is the only guardrail, your system is lazy. You want automated guardrails first. Tests, policy checks, confidence scoring, rollback paths, audit trails, blast-radius limits. Then, after all that, maybe a human. JOSH: So the tired part is the volume? ERIK: The volume, the context switching, and the fake importance. A person gets pinged, opens a ticket, reads six logs, figures out the model was unsure about a field name, clicks approve, and goes back to their real work. Do that fifty times and they'll approve garbage just to make it stop. JOSH: That's wild, because the company probably thinks it added safety. ERIK: Exactly. It added theater. Safety is when the system blocks the bad action before it becomes a human's problem. [beat] ERIK: PrimeBus auto-merger has run 2295 attempts since June 5. 1507 merged, 788 blocked by Gandalf, 0 escalated to me. That's the story. Not that every agent is perfect. The story is that Gandalf catches the stuff that shouldn't ship. JOSH: Wait, zero escalated to you? ERIK: Zero. And that does not mean zero risk. It means the pipeline has layers. The agent proposes a fix. Tests run. Review logic checks it. Gandalf blocks suspicious changes. PrimeBus emits telemetry. If it can't prove the change, it doesn't merge. JOSH: So you're not sitting there approving tiny changes all day. ERIK: No. That would defeat the point. If I have to approve every small fix, I don't have automation. I have a very expensive intern with a modal dialog. [beat] JOSH: Where do humans belong then? ERIK: Policy, exceptions, and intent. Humans should decide the rules of the factory. What counts as acceptable risk. What systems are off-limits. What requires rollback. What kind of customer impact is allowed. But once the rule is clear, the machine should run it. JOSH: Give me a real example. ERIK: Say PrimeRestorer sees a backup failure. The agent doesn't need to ask me if backups matter. That's already policy. It can inspect the job, compare prior runs, check S3 offsite status, retry if the failure is transient, and open a real incident only if recovery risk remains. JOSH: And if it can't tell? ERIK: Then it should say exactly what it knows and exactly what it doesn't. Not, hey Erik, something happened. That's garbage. It should say, database backup failed, local snapshot exists, offsite copy is stale, retry failed twice, suspected credential rotation. Then I'm useful. [pause] JOSH: Next deep dive. LM Studio Bionic. An AI agent for open models. Why does this matter if Claude and GPT are already so strong? ERIK: Because control matters. Hosted models are great. I use Claude constantly. But not every workload belongs in somebody else's cloud. Local agents give you privacy, predictable cost, and the ability to run close to the files, logs, and tools. JOSH: But are open models good enough now? ERIK: For a lot of work, yes. Not everything. I wouldn't hand every architecture decision to a small local model. But classification, extraction, log triage, draft fixes, test repair, simple code edits, document search. Absolutely. [beat] ERIK: The trick is routing. Don't ask one model to do everything. PrimeRouter exists because model selection should be a systems problem, not a vibes problem. JOSH: How do you decide what goes where? ERIK: Cheap model first for low-risk work. Stronger model when the task has ambiguity. Hosted frontier model when the outcome matters and the local model can't prove it. And if providers are down, fail over. If latency spikes, route around it. Same way network engineers have always thought about paths. JOSH: That's very networking-brained. ERIK: Yeah, because it works. People keep treating AI like a magic chat box. It's infrastructure. It has routing, health checks, retry logic, telemetry, budgets, and ugly edge cases. JOSH: So LM Studio Bionic is interesting because it puts the agent closer to the user? ERIK: Right. The desktop becomes a worker node. Local files, local scripts, local models. That matters for small teams. It also matters for companies with rules around data. You can have an agent inspect a private repo or a log bundle without sending the whole thing out. [beat] JOSH: What's the catch? ERIK: Local agents need discipline. They can still make bad edits. They can still loop. They can still chew through time doing nonsense. You need permissions, sandboxing, dry runs, and a clean way to review diffs. JOSH: That's the same lesson again. ERIK: It is. Agents are not the product. The control plane is the product. [beat] ERIK: In my setup, 12 agents are in the Bobaverse fleet right now. Neo, Homer, Bill, Echo, Gandalf, Claude plus GPT. That only works because they talk through PrimeBus and telemetry. If they were all just random chat windows, I'd have chaos with invoices. JOSH: Chaos with invoices is a strong enterprise product category. ERIK: It already exists. People call it AI transformation. [beat] JOSH: If somebody listening wants to try local agents this weekend, what should they build? ERIK: Build a log triage agent. Very small. Give it one directory of logs, one known error pattern, and one output format. It should summarize what changed, identify likely cause, and create a suggested next step. No write access at first. JOSH: Start read-only. ERIK: Exactly. Read-only first. Then allow it to write a report. Then allow it to open a ticket. Then allow it to run one approved command. That's how you build trust. You don't start by giving it prod credentials and a dream. [pause] JOSH: Third story. Kimi K3. Open frontier intelligence, 2.8 trillion parameters, 1 million token context window. Is bigger still better? ERIK: Bigger is useful, but bigger is not the whole story. The 1 million token context window is the part people will feel first. Long repo analysis. Giant policy docs. Whole incident timelines. That changes how you package work for a model. JOSH: But do people actually use the long context well? ERIK: Usually no. They dump everything in and ask for magic. Long context is not a trash can. It's a workspace. JOSH: What's the difference? ERIK: A trash can is every log, every doc, every issue, every half-related Slack thread. A workspace has structure. Here's the goal. Here's the repo map. Here are the failing tests. Here are the constraints. Here are the files that matter. Here's what you must not touch. [beat] ERIK: Long context gives the model room to reason, but you still have to brief it like an engineer. JOSH: How does that compare to ScanBrief? ERIK: ScanBrief scored 93 items across 54 sources today. It doesn't just throw links at me. It scores, ranks, dedupes, and gives me the signals. Same idea. Context has to be filtered before it's useful. JOSH: So the model window got bigger, but the input still needs taste. ERIK: Taste, policy, and structure. If you feed a model noise, you get a confident summary of noise. That's not intelligence. That's a well-written junk drawer. [beat] JOSH: Where would you use a million-token model in your own systems? ERIK: Codebase review. Incident review. Contract review. Anything where the hard part is cross-file or cross-document reasoning. GovContractScanner could hand over an RFP, prior awards, my service catalog, and pricing constraints in one shot. Then the model can say, this is a fit, this is not, this clause is risky, here's the response outline. JOSH: That sounds expensive. ERIK: It can be. That's why routing matters again. Don't use the giant model to classify a subject line. Use it when losing context causes mistakes. JOSH: Are open frontier models a threat to the big labs? ERIK: Yes, but not in the cartoon way. The big labs still have the best models for many hard tasks. But open models keep pushing down the cost of good-enough intelligence. That means more automation runs locally, more businesses build private agents, and hosted models get used where they truly earn the bill. [beat] JOSH: Where does Decoy Font fit into this? The one that hides text from AI systems. ERIK: That's a fun one because it's basically adversarial design for vision models. It makes text readable to humans while confusing AI that processes the pixels differently. JOSH: Useful or gimmick? ERIK: Both. As a product, maybe niche. As a signal, important. People are going to start designing content, documents, and interfaces that treat AI readers as a separate audience. JOSH: Like robots are now part of accessibility testing. ERIK: Exactly, except sometimes you want the robot blocked. Domestic abuse safety guides, sensitive screenshots, private notes, internal docs. GrapheneOS being recommended for domestic abuse victims was also on the list today. That overlaps. Privacy is not a nerd preference. Sometimes it's physical safety. [pause] JOSH: That makes the AI text detection headline feel a little weird too. Classical machine learning for detecting LLM-generated text. ERIK: Detection is messy. Classical ML can catch patterns, but the target keeps moving. The better question is not, was this generated by AI. The better question is, can you trust the source, can you verify the claims, and who is accountable. JOSH: So you don't care if AI wrote it? ERIK: I care if it's wrong. I care if nobody checked it. I care if it's being passed off as expertise when it's just autocomplete with confidence. [beat] ERIK: AI writing is not the enemy. Unowned output is the enemy. If a system sends a customer email, opens a ticket, edits code, or trades based on a webhook, somebody designed that process. Own it. JOSH: PrimeTrader fits that, right? ERIK: Yep. PrimeTrader can react to TradingView webhooks, but the guardrails matter more than the signal. Bad automation in finance is not cute. It costs money. So the system needs limits, logs, and a kill switch. Boring stuff. The stuff that keeps you from learning expensive lessons. [pause] ERIK: This episode is sponsored by Prime Automation Solutions. If you're still doing it manually, we automate it. Also, special on a website — $250. primeautomationsolutions.com [pause] JOSH: Alright, what's the AI pro tip today? ERIK: Make your agent produce evidence before it produces an answer. [beat] ERIK: Don't ask, fix this test. Ask, inspect the failure, name the failing assertion, identify the smallest suspected cause, propose two fix options, then apply the lowest-risk patch. After that, run the exact test and report the command and result. JOSH: That's more work up front. ERIK: Yes. That's the point. Lazy prompts create lazy agents. You want the model to leave tracks. If it can't explain what failed, what it changed, and how it verified the fix, it doesn't get to touch the next thing. [beat] ERIK: This works for code, Terraform, Kubernetes manifests, NSO service packages, even Selenium test repairs. Evidence first. Action second. That's your tip. Use it. [pause] JOSH: Binge all five episodes this weekend plus our YouTube shorts — links at buildorbereplaced.dev. [pause] JOSH: One more thing — we started a Discord for builders. If you're shipping AI, automation, or anything that makes a human obsolete — come hang out. Link at buildorbereplaced.dev. ERIK: Post what you built. We'll post what we're building. Real wins, real builds, no fluff. [pause] ERIK: Build or be replaced. JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.