Principle · SocialWhen walking past a coffee shop with a woman you're talking to, ask if she's down to go in for a coffee, framing it as something you were already going to do yourself (e.g., 'I was actually thinking about going and getting a coffee or a tea. You wanna go hit up Starbucks?').
Coach Kyle · From Hi to Coffee Date (LIVE DEMO)
Creator perspective · MoneyRene Haas argues AI tooling is highly effective for the verification, validation, debug, and documentation stages that dominate chip design timelines.
No Priors: Artificial Intelligence | Technology | Startups · Redefining Chip Architecture with Arm CEO Rene Haas
Creator perspective · MoneyRene Haas explains that AI tools are still not mature in two areas: RTL generation and physical design/implementation at best of class. His explanation is that the models are trained on what is available publicly, and a lot of chip design information is quite proprietary. He sees massive opportunity for the ecosystem and everyone in the industry to make these tools better, and believes it is only going to get better.
No Priors: Artificial Intelligence | Technology | Startups · Redefining Chip Architecture with Arm CEO Rene Haas
Creator perspective · MoneyKent C. Dodds argues that engineers who just take a ticket and turn it into an implementation look replaceable, and that lacking product sense or design sense makes you easy to replace as agents get competent.
AI Engineer · Build the Right Thing: Product Engineering (Part 1) — Kent C. Dodds, EpicProduct.engineer
Report · MoneyIn an AI-generated demo call, the user asked for a refund for order 14, the agent misheard and started a refund for order 40, the user corrected it, and the agent appeared to confirm the correction — but the confirmation actually interrupted the user and the refund stayed on order 40. The transcript looked fine while the actual call was broken. There was also 2.4 seconds of dead air before the agent responded, a latency spike after the initial refund request, and the agent talked over the caller during the correction turn.
AI Engineer · The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize
Report · MoneyYou can create evaluators either in-product via Alex (Arize's in-product agent) or via skills in your coding agent; either way the evaluator is created for you and runs automatically. However, Signal does not automatically create its own evaluators right now, because evaluators take time and money to run — you get the suggestion and then create the evaluation yourself.
AI Engineer · From Vibes to Production: Evaluating and Shipping AI Agents That Work 201 — Laurie Voss, Arize AI
Report · MoneyClosing the observability loop progresses through four stages: no observability, then observability, then automatic problem detection, then automatic problem fixing. A monitoring agent like Signal continuously reads traces, finds wrong patterns, and suggests fixes without manual asking.
AI Engineer · From Vibes to Production: Evaluating and Shipping AI Agents That Work 201 — Laurie Voss, Arize AI
Report · MoneyThe joy of agentic coding, per Voss: you never touch the code — you say 'look for problems,' it finds one, you say 'fix that problem in particular,' and it does. The full loop was demonstrated: from no observability, to observability and knowing what's going on, to finding problems automatically, to fixing problems automatically — the loop is closed. After the fix, querying 'Show me all toys under $9' returns the correct results, and the traces show the input and output including all those toys under $9.00.
AI Engineer · From Vibes to Production: Evaluating and Shipping AI Agents That Work 201 — Laurie Voss, Arize AI
Creator perspective · MoneyFrye distinguishes bug ownership: application-level bugs from models misbehaving are the app developer's problem (though they show up on your dashboard), while model quality bugs are the engine's problem — verify you are doing inference correctly and that performance optimizations didn't change model behavior. One of the most common model quality bugs occurs at model release: tokenizers are often slightly bugged because they both tokenize and apply chat templates. These are typically beaten down within a couple weeks after release, but you always have to be on guard. His hot tip: log the token IDs — set your engine up so token IDs are logged.
AI Engineer · What Is an Inference Engine, Anyway? — Charles Frye, Modal
Creator perspective · MoneyPromoting shared memory makes cost go down tremendously: the first time you burn lots of tokens going back and forth with the LLM, but the next time another agent or session retrieves it from Mako memory with far fewer tokens. The alternative is keeping things in empty files or constantly traversing with every agent's LLM, burning through tokens again and again to retrieve. The difference with Mako is just an MCP tool call.
AI Engineer · Agent Memory Is Solved. Agent Learning Isn't. — Karthik Ranganathan, Yugabyte
Report · MoneyThe host claims that a lot of times a turnaround can be achieved not by increasing activity or getting people better on the phones, but by removing waste in prospecting. He suggests categorizing prospect personas into A, B, and C tiers, and replacing C tier prospects with A tier prospects without increasing activity to get better meetings.
30 Minutes to President's Club | No-Nonsense Sales · #606 - How To Build An Elite Outbound Sales Team (Get Reps Dialing Fast)
Report · MoneyTo manage many sessions: Fleet (nicknisi/fleet on GitHub) lists agents by project and session name with auto-generated summaries and completion notifications (use a visual indicator rather than sound when running 12+ agents); the Sessions tool installs via 'brew install sessions'; drive everything through TMUX; for fewer than 10 tabs, just rename Claude sessions with /rename and rename terminal tabs; turn on Claude's recap feature for an italicized summary when refocusing a window; session summaries live in JSONL files so you can write your own tool.
AI Engineer · Lifestyles of the AI-Native — Nick Nisi & Zack Proser, WorkOS
Creator perspective · MoneyFor frontend problems, use Claude Design first — a specific instance of Claude adapted for design that develops a design system persisting across sessions and reusable — with higher spend up front, similar to ideation. Inject adversarial review via hooks toward the end of an implementation loop, loading skills designed to test UX and front-end best practices; make the adversarial reviewer as negative, persnickety, and nitpicking as possible — have it fail the first 25 builds saying 'this is not good enough. Keep going. Fix these things.' Once upfront planning and adversarial review are in place, you can engineer a loop where you make coffee and come back ~40 minutes later to work that's close to merging and you're actually happy with it.
AI Engineer · Lifestyles of the AI-Native — Nick Nisi & Zack Proser, WorkOS
Creator perspective · MoneyUse the custom Nick Nisi ideation plugin (in the cloned repo's cloud settings) instead of built-in ones — it specifically breaks work up into manageable specs that can run in independent context windows, to manage context size and avoid bloat. Break work into small atomic pieces that run in independent context windows.
AI Engineer · Lifestyles of the AI-Native — Nick Nisi & Zack Proser, WorkOS
Creator perspective · MoneyWhen agent-generated PRs arrive in high volume (stacks of 30 PRs, sometimes 50,000 files in a single PR), treat review as a large system complexity problem. Use tools like Devin and Greptile for first-pass analysis; treat a Greptile 'five out of five ready to merge' score as just a first signal because the tools are wrong sometimes. Fail out PRs scored twos and ones and kick them back; for fours and fives, tell agents in a loop to babysit the PR and fix comments until there are no more comments; near the end, use your human brain to spot-check sensitive files. Zack is building Dab, a tool that tells you a story about your PRs so you can stay in the loop.
AI Engineer · Lifestyles of the AI-Native — Nick Nisi & Zack Proser, WorkOS
Report · MoneyIn the demo, the speaker used GPT-3.5-turbo as the judge model to show some of its biases. The judge prompt should define the model as a helpful assistant that judges answers to a given scenario, and ask the model to respond with only the number of the correct answer, no punctuation. The judge prompt should include the scenario input and the agent's answer, each on its own line, without the human verdict. Expect the judge's output to match the scenario's human verdict — that's the agreement check. LLM judges are non-deterministic, so you should run the same eval multiple times to see the real behavior. LLM judge calls can fail due to timeouts, so add a timeout (e.g., 30 seconds) to handle this. The speaker uses Cursor to write code and stores environment configuration in a .env file with
AI Engineer · Evals in AI: A Deep Dive — Tejas Kumar, IBM
Report · MoneyUnlike the coding use case, applying agents to production debugging involves a lot of human-agent back-and-forth: the agent confidently claims a diagnosis (e.g., a memory leak from a 500-error spike, or the readiness probe), and the human must check events to see whether it actually happened, iterating until the real problem is found. Even when the agent does much of the work, the human remains on the hook to iterate and second-guess it.
AI Engineer · Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Principle · MoneyThe other big conclusion of the report is just how much of a power-user game consumer AI is right now — a small share of heavy spenders drives most revenue, so builders should optimize for the top cohort rather than average users. This is the speaker's actionable takeaway from the spend-concentration data. Today's heavy AI spenders ($900/month power users) should be treated as harbingers of mainstream behavior as more people feel empowered to create.
Interpretation of the report's own data.
The a16z Show · The Top 100 Consumer AI Apps: Who’s Actually Paying?
ReportGoogle open-sourced a new Gemini API CLI that you can use to get started working with Gemini models and test the models locally using simple CLI commands. The CLI lets you iterate on your coding agent and anti-gravity agent locally, and once ready, it packages it as a single folder, uploads it, and creates an agent for you.
AI Engineer · An Interaction Is All You Need — Ivan Leo, Google DeepMind
ReportTo analyze a GitHub repository, pass it in an API call to the anti-gravity agent, which boots up a remote sandbox and starts investigating the repository. The agent calls list files to get what's in the workspace, reads each individual file, and then generates a final report. The agent can research, reason, and handle repository analysis autonomously in a remote sandbox without any intervention from the user's side.
AI Engineer · An Interaction Is All You Need — Ivan Leo, Google DeepMind