Adil Wali

If AI Does the Junior Work, Who Becomes the Expert?

Adil Wali — Mon, 19 Oct 2026 13:15:00 GMT

AI agents are starting to take over the work people find pointless. Davide Ghia on why a lot of that work is how juniors learn the job, and who will be senior enough to supervise the agents.

Chapters

Timed transcript · Full show notes and sources

Sources

Will AI Agents Free Us From Meaningless Work? A Human-Centered Analysis (Ghia, Ranjit, Cerquitelli and Quercia) — arXiv

Disclosure

Adil Wali invests in startups. At 20:02 he describes how he looks for companies to invest in. At 21:12 and 31:09 he talks about OpenAI and Anthropic. Nothing in this episode is investment advice.

Inference Can Wait. The Grid Can't

Adil Wali — Tue, 13 Oct 2026 13:00:00 GMT

The grid doesn’t have an energy problem; it has a peak-power problem. Chris Dunlap on why AI data centers face a choice between connecting now and flexing in the worst hours, or demanding firm power and waiting years—and which AI work can actually wait.

Chapters

Timed transcript · Full show notes and sources

Sources

Quantifying AI data center flexibility as a resource adequacy asset — Research Square (preprint)
Quantifying AI data center flexibility as a resource adequacy asset (working paper) — The Institute for Climate and Sustainable Growth, University of Chicago

Disclosure

Adil Wali invests in startups. At 50:53 he mentions that startups he has invested in are moving workloads to Chinese open-weight models. Nothing in this episode is investment advice.

I stopped building my AI chief of staff. Six weeks later, a startup in the same space raised at $10B.

Adil Wali — Wed, 07 Oct 2026 18:17:24 GMT

Earlier this year, I built Chief, an AI chief of staff for myself. By August, I felt like I was on a collision course with Anthropic, OpenAI, and Meta, so I pulled the plug. Then Instinct, the viral AI assistant, raised $1B at a $10B valuation, and I was left scratching my head. Did I make a mistake when I stopped actively building Chief?

Why Chief

I’ve had an executive assistant for most of my professional career and I rely on them heavily. When I left Klaviyo, it didn’t make sense to put someone on my personal payroll when theoretically I finally had time to manage my own email and calendar.

I still didn’t want to do those things, though :)

So I set out to build an assistant that could know my preferences, my priorities, and keep me on track. I imagined a proactive employee that would have a daily morning brief, summarize my most important emails, draft background notes for meetings that day, and understand the key people in my life.

I built it, and it worked. It actually worked exceptionally well.

I had an intense amount of fun iterating on it. I am highly technical, but I wouldn’t have been able to build this on my own until AI came along, which made solving these problems even more interesting.

Chief booked restaurant reservations for me, prepared briefing notes for upcoming meetings, and texted with my friends and family. It summarized and processed my inbox and highlighted what I needed to focus on. It even cajoled me when I didn’t follow up on things I said I would.

The core of Chief was ultimately a memory system that was built for AI to use rather than humans. The belief was (and still is) that the best UI for a personal database is an AI agent. It just needs the right tooling under the hood.

Why I Pulled the Plug

So why pull the plug? First, as cool and useful as it was, it felt like a second door to the agents in my life and it was competing with a primary door (Codex and Claude Code) that was so ingrained in my routine that it would be tough to displace it. I concluded: If that was my case, it would be everyone else’s.

Second, it was clearly on a collision course with what the frontier labs were doing. I built a Chrome plugin early in the Chief journey so it could drive the browser for me. Then I built a credential/secret storage system. Then I built connectivity to iMessage and WhatsApp. And at every step, it felt obvious and clear to me that the frontier labs would ship the same things (which they mostly have done).

We can’t forget that Anthropic and OpenAI own the models, have unlimited funding, and can subsidize any product they want to for as long as they want to. As they started to connect to people’s email, calendars, and documents, I predicted they would quickly come for this problem space and build it into their own products, making a standalone tool no longer a viable business.

It was also very expensive to run, which would be fine if it were just a productivity tool for me. The expensive part wasn’t talking to it. It was everything Chief did before I asked: reading everything that came in, keeping its memory current, and constantly deciding what was worth flagging. In July, Chief was spending about $1,600 a month per user on models. Only about $10 of that went to the conversations where someone actually asked it for something.

Model prices are falling fast, and I was already moving to cheaper models. But every new user would bring a whole inbox, Google Drive, and set of notes to read and remember, whether or not they ever opened the app. As I thought about productizing it, I couldn’t see that scaling without a funding round, and I’m explicitly trying NOT to commit to something right now. Money creates commitment and expectations. It locks you in.

So in August, I killed Chief as it was originally scoped. I didn’t throw it all away, though. I dropped the proactive assistant and kept the memory underneath it. Instead of building a second door, I put Chief behind the one I already use every day. Today, Chief is the memory my AI agents work from, and it’s how I’m looking around corners at where agents are headed. (More on this in another post.)

Do I regret it?

Instinct’s round made me stop and think. Chief was more powerful and capable than Instinct (not to take anything away from them, I was solving a deeper problem for an executive audience), and only weeks earlier I had decided it had no future. I wondered if I was underselling myself or the product. Maybe I hadn’t thought enough about the economics or the path to market. Maybe I wasn’t creative enough about ways to cut cost. Maybe I was focused too much on building it for myself as a customer and not enough for a mass market.

Three weeks before that $1B round, Meta launched Muse. And having used it, I’m convinced I was right all along. Two reasons. First, how functional it is, despite being built for the everyday user. It’s not that Meta does more than Chief did (it doesn’t). It’s that it does enough. And it does it reasonably well.

Meta got a lot right with Muse -- it’s proactive, but not pushy. It’s relatively smart, doesn’t hallucinate much, and makes all of the third-party connectivity simple and effective. Putting a VM behind every Muse agent was a stroke of genius and the kind of move that’s nearly impossible for anyone to compete with, because a dedicated computer for every user, mostly for free, only works if you can afford to subsidize it. I’ve used Muse to book concert tickets, to monitor prices for sporting events, and to help me plan trips, all with almost no hassle or issue.

I knew they were really cooking with gas when I saw the secure credential storage system (I built one too), and I figured that most simplified agents wouldn’t go this far, but they did.

And the day after Instinct’s round, OpenAI launched dots: an always-on agent with its own cloud computer that does “proactive research” in the background, included in a ChatGPT Pro plan at no extra cost. That background work was almost all of the $1,600 a month per user Chief was spending.

The second reason, and maybe the more important one: building a money-losing product that you subsidize is a game for those with very deep pockets. And I continue to believe that the companies best positioned to do that are the ones that are able to spend from their cashflow as opposed to those who have to take in an infusion of inorganic money to get there.

Instinct’s investors clearly see something, and I hope they’re right. But Instinct’s functionality seems less complete than Muse’s, it doesn’t have an app yet, and it is now competing with a company that has 3.6 billion daily users and gives Muse away free for most of what people need. Perhaps most importantly, Meta can subsidize it indefinitely if it chooses to. A hype cycle can carry a valuation for a while. When it ends, you actually need to have a business. Time will tell if they do or not.

So, from one founder to others out there: don’t discount the power of incumbents when you’re exploring a new idea. Don’t commit to a path too soon, and don’t be afraid to acknowledge that a moat you thought you had is eroding.

If I hadn’t, I might have sunk more time and money into an idea that was promptly swallowed by one of the biggest fish in the sea. Instead, I’ll put that time into learning, talking to smart people, and keeping my eye out for where the real, defensible opportunity is for my next thing.

And as much as it might sting to have shelved a product that apparently there is a ton of investor appetite for, I have to remind myself that raising money isn’t winning. Building a successful business is winning. No regrets.

We’re Entering the Era of Private Software — Maxim Khailo

Adil Wali — Thu, 03 Sep 2026 23:49:54 GMT

Watch the full conversation on YouTube, or listen in the player above.

Open source you share. Proprietary you buy. Maxim Khailo argues the next era is software you can’t buy at all—and states the one condition that would end it. Plus why a deployed model doesn’t learn, and the energy bill under every agent fleet.

Chapters

Timed transcript · Full show notes and sources

Disclosure

Adil Wali is an investor in Thetix. Maxim Khailo is associated with the company, and it is discussed in the episode from 31:49.

Startups aren't getting cheaper to build.

Adil Wali — Wed, 02 Sep 2026 14:15:42 GMT

US venture capital just had its biggest half-year on record: $412.7 billion deployed in the first six months of 2026. Deals under $100 million took 12.5% of it — $51.4 billion, split across every US venture deal below that threshold.

Inside that same record half-year, the median seed round went the other way. $3.0 million so far in 2026, down from $3.5 million across 2025 — the only stage where the median fell while every other series rose.

A record amount of money invested but shrinking seed rounds point to the same thing. The money is stacking up at the top.

Figure 1 — Four checks, and everyone else. Every note is $10B, drawn at the same scale on every row. The top three rows are individual disclosed rounds. The bottom row is every US venture deal below $100 million in the same half, added together. The rows aren’t additive — those mega-rounds sit inside the $412.7B too. Source: PitchBook–NVCA Venture Monitor, H1 2026; individual round sizes as reported by PitchBook and Crunchbase.

Put the individual checks side by side and the shape stops being abstract. OpenAI raised $122 billion in a single round. Anthropic raised $95 billion across two. xAI raised $20 billion. Every US venture deal under $100 million — every seed, every Series A, every small round cut in the country over six months — came to $51.4 billion between them.

One company’s round was worth more than twice that entire market.

That behavior is consistent with a narrative I hear everywhere: how AI has made starting a company so cheap. After all, if building a company is cheap now, startups should need less money.

There are two unspoken assumptions in that assertion worth challenging. First, that building a product is the same as building a company. Second, that all other costs aside from product development also stay the same or go down.

I don’t dispute that building software is cheaper with AI. The other assumptions, however, don’t hold up for me.

The question nobody seems to be asking is: what are the second- and third-order effects of making software dramatically cheaper and easier to build? Ignoring that line of thinking is likely to lead a lot of founders — and a fair number of investors — down the wrong path.

Subscribe free to keep reading

Get the rest of this essay—and future field notes on building companies in the AI era—delivered to your inbox.

Subscribe now

AI Makes Engineers 100x Faster. Why Aren’t Companies? — Jared Green

Adil Wali — Wed, 19 Aug 2026 23:08:33 GMT

Watch the full conversation on YouTube, or listen in the player above.

AI can make one engineer feel 100x faster. Jared Green explains why companies aren’t becoming 100x more effective—and why the bottleneck moves into review, coordination, technical debt, and the systems around abundant code.

Chapters

Timed transcript · Full show notes and sources

Disclosure

Adil Wali is an investor in Autar. Jared Green is Autar’s founder, and the company is discussed in the episode.

SaaS is Dead; All-In on AI?

Adil Wali — Thu, 13 Aug 2026 23:45:00 GMT

Watch the full episode on YouTube, or listen in the player above.

The market erased roughly $272 billion from ten durable SaaS companies while assigning more than $1 trillion to frontier AI labs. I explain why that trade may be underestimating systems of record, authority, and operational moats.

Chapters

Timed transcript · Full show notes and sources

Sources

Salesforce FY2026 Form 10-K — U.S. Securities and Exchange Commission: Salesforce reported trailing attrition of approximately 8%; the metric excludes Slack self-service and current-year acquisitions.
ServiceNow 2025 Form 10-K — U.S. Securities and Exchange Commission: ServiceNow reported a 98% ACV-based renewal rate. Its definition can count ordinary downsells as renewed.
Workday FY2026 Form 10-K — U.S. Securities and Exchange Commission: Workday reported approximately 97% gross revenue retention, excluding expansion.
HubSpot 2025 Form 10-K — U.S. Securities and Exchange Commission: HubSpot reported approximately 88% gross customer-dollar retention and changed its treatment of partner commissions in 2025.
Workiva 2025 Form 10-K — U.S. Securities and Exchange Commission: Workiva reported 97.2% gross revenue retention and 112.8% net revenue retention for 2025.
New sales agents in Microsoft 365 Copilot — Microsoft: A concrete example of a seller working through an agent while Dynamics 365 or Salesforce remains the underlying record.
HubSpot agents move to outcome pricing — HubSpot: Customer Agent and Prospecting Agent are priced by resolved conversation or recommended lead rather than only by access.
Salesforce FY2026 fourth-quarter results — Salesforce: Early Agentforce bookings were primarily expansion within Salesforce's installed base.
Workday FY2026 investor presentation — Workday: Scale and embeddedness evidence for the ERP and HCM discussion.
Claude Cowork research preview — Simon Willison: January 12, 2026 is the base date used for the market comparison. The episode says “since,” not “because of.”

Disclosure

This episode reflects my personal analysis and investment views. I own and may buy or sell positions in public and private technology companies discussed here. The market comparison uses different pricing mechanisms and valuation dates for public companies and private AI labs, so it is directional rather than a synchronized index. This is not investment advice.

I Spent 471 Million AI Tokens Rebuilding a Startup I Invested In

Adil Wali — Wed, 05 Aug 2026 04:20:00 GMT

Watch the full episode on YouTube, or listen in the player above.

I rebuilt the product of a startup I invested in, then measured exactly what 471 million processed tokens, 64 prompts, and a week of judgment actually bought.

Chapters

Timed transcript · Full show notes and sources

Sources

The S&P 500 and Trump Accounts — S&P Dow Jones Indices: S&P 500 total return figures through December 31, 2025.
Nasdaq-100 Total Return Index fact sheet — Nasdaq: Ten-year return through June 30, 2026; a different window from the S&P figure.
World Wealth Report 2026 — Capgemini: U.S. high-net-worth population as defined by investable assets.
Survey of Consumer Finances — Federal Reserve: Direct ownership of publicly traded stock among U.S. families.

Disclosure

I’m an investor in Thetix. That’s part of why I care enough to go this deep. The redesign shown is an independent exploration, not an announced Thetix feature. Nothing here is investment advice; every portfolio example is product analysis.