Summary

Ali Ghodsi’s central claim: “I think we have AGI. We really have it” — by the definition his 2009 Berkeley AMP Lab used, it’s already satisfied, and the industry is just “moving the goalpost.” He sorts the field into three camps: the superintelligence quest (frontier labs, most of the capital, “I would be very worried there”), the Turing-Award researchers (Sutton, LeCun — sober, 20 years out, “probably the ones that are right, unfortunately”), and camp three — Databricks and Glean — extracting economic value from the AGI we already have.
“The LLM is a commodity” — interchangeable like gas stations, “just compare price,” with users switching models in a day unlike any prior platform battle. Model companies can still be valuable (“TSMC is very valuable”) but as fabs; the real moat is proprietary data and business process — “there’s not an AI out there that understands your secret sauce and your data. That’s not a commodity.”
On the capex math — ~$250B to Nvidia implying ~$500B capex needing ~$1T of AI revenue vs a $400B total software industry — Arvind Jain’s resolution: AI isn’t extending software, it’s converting services dollars, an industry “25 times larger than software.” Ali’s answer is camp-dependent: if superintelligence lands, “any of your cost equations pale in comparison”; camp three doesn’t need it.
Is there a bubble? Yes, but not binary: “there are startups with zero revenue worth 10, 20, 30 billion. That’s a bubble.” Yet both call OpenAI and Anthropic up over 12 months — ChatGPT and Gemini “on fire,” coding having “only eaten into a small portion of that market.”
Ali reframes the MIT 95%-failure stat as healthy: “that’s actually what you want” from an experimentation phase, and hopes for similar stats next year. The working 5%: RBC agents producing equity research notes 15 minutes after an earnings call vs a 2-hour industry standard, Merck’s “Teddy” transformer for gene-regulatory drug discovery, and 7-Eleven’s fully agent-automated marketing stack.
Value accrual call: Arvind thinks the intelligence layer stays thick — “maybe half of enterprise value” — while Ali says most value goes to apps, “I just don’t know which apps”, invoking 1998: everyone bet on Cisco routers and portals, the winners were Facebook/Airbnb/Uber. Software isn’t dead (Salesforce is “a full ecosystem of workflows,” not a database), but data entry is the wedge — “Zoom is really the perfect data entry application.”
Longs and shorts: Ali is long agents and speech (“as long as you’re using a keyboard, we haven’t nailed speech” — keyboards “basically going to disappear”), Arvind calls coding and customer-service automation “a little bit over hyped.” Brad is long proactive AI that comes to the user — the shift that takes “5% power users to 100%.” Glean, fresh off a $200M revenue run rate, is building toward a privileged personal work companion.

Deep dive

1. 95% of AI projects failing is “actually what you want”

Ali’s read of the MIT report is contrarian and unhedged: “if all of your projects are failing, that means you’re trying enough… when I read the study it was not a surprise for me” — and he hopes for similar stats next year, because the phase rewards aggressive experimentation, not hit rate.
The 5% specimens, one per vertical: Royal Bank of Canada’s agent ingests the earnings report, prior quarters, competitor filings and market news, and ships a full equity research note 15 minutes after the call vs a 2-hour industry standard. Merck’s “Teddy” (transformer-enabled drug discovery) predicts which genome is missing when one is removed — “it really understands the gene regulatory network.” 7-Eleven runs an agent-automated marketing stack; Ali thinks “the marketing stack is going to get disrupted pretty heavily” because fine-grained segment-level content creation was previously manual labor.
The caveat kept intact: “it’s not just you can unleash the agent and it just works — it’s an engineering art” needing evals, productionization and a great team. And honestly: “even with Databricks, we’re not just the 5% — we have some of that 95% too.”

2. “The LLM is a commodity” — the moat is data your competitor doesn’t have

Ali’s econ-class framing: commodity means interchangeable — “you can get gas from this gas station, you can get gas from that gas station… just compare price.” People switch LLMs in a day, unlike iPhone/Android or “Google Sheets versus Excel — a huge religious battle inside our company.” Commodity doesn’t mean worthless — “TSMC is very valuable” — but the labs become “fab-like companies.”
What isn’t commodity: “there’s not an AI out there that understands all your business processes, your secret sauce, and your data.” The industry’s twin failure modes: companies building “commodity stuff” any competitor can replicate, and “a lot of demo-ware — it’s really easy to make cool demos with an AI.”
The host’s Altimeter house line, endorsed by Ali: “your AI strategy starts with your data strategy.”

3. Not RPA redux — but the frozen-model problem is unsolved

Both CEOs dismiss the “same movie, bigger budgets, better actors” comparison. Ali: RPA was “rule-based… zero learning,” brittle to anything unexpected, versus a learning agentic system that generalizes. Arvind won’t even engage: “I would not compare these two technologies at all” — first seeing AI “was basically magic.”
Ali’s honest caveat — several high-profile genAI-replaces-RPA startups have already failed because of today’s paradigm: “you bake a model… and then you freeze it.” What’s needed is AI that keeps learning while clicking around a desktop, and “we haven’t really nailed computer use yet.”
Both offer their own 95%: Glean’s fine-tuning work “didn’t really pan out”; Arvind still can’t get his weekly-priorities rollup agent working despite the AI having “all the context” — “AI is just one more tool in the toolkit.” Databricks’ first pass at automating software engineering failed too: “nothing wrong with the AI. The problem is the humans and how we were organized.”
Advice to CIOs planning budgets: “the winners are yet to be identified” — experiment with more vendors, shorter-term contracts, and pick products testable quickly, not 6-month implementations.

4. Three camps resolve the trillion-dollar physics problem: “we already have AGI”

The host’s setup: ~$250B to Nvidia implies ~$500B of capex needing roughly $1T of AI revenue — versus the entire software industry at $400B. Arvind’s resolution: AI is “not extending software in a marginal way,” it’s grabbing revenue from the services industry, “25 times larger than software” — those service dollars are converting into AI dollars.
Ali’s taxonomy: camp one, the superintelligence quest — frontier labs, scaling-laws mentality (“whoever has the most GPUs and the most data wins”), pointing to Olympiad benchmarks, promising recursive self-improvement, cured cancer and 10x GDP — “so what the hell are you talking about that there’s a physics problem?” Camp two, the Turing-Award scientists (Rich Sutton, Yann LeCun): autoregressive next-token prediction “is not how humans learn” — no child reads the internet four times before speaking — and the real thing is 20 years out. Camp three: Ali and Arvind.
The goalpost argument: “I think we have AGI. We really have it… it’s a false premise to start with.” By the definition used at Berkeley’s AMP Lab in 2009, AGI is satisfied — he went back and asked, and colleagues agreed “we’ve changed the definition now.” The mission: “expand that 5% to 10, 20, 30%… we have the AGI we need. Let us just do our engineering.”

5. Value accrues to apps — “but I just don’t know which apps”

A live disagreement: Arvind guesses the intelligence layer stays “pretty thick — maybe it’ll capture half of the enterprise value,” while Ali relegates models to fabs and elevates data plus the governance/security layer (“What if it’s using a Chinese model? Oh, here’s the provost’s salary information — oops”) — but concludes “most of the value will accrue to the apps.”
The 2000 analogy told in full: in ‘98 we thought it was Cisco routers and portals “with a hundred links”; the winners turned out to be Facebook, Airbnb, Uber. But it’s not binary for incumbents — Amazon and Google already existed in ‘98 (Google “only a $300 [million, likely] company”) — so Databricks and Glean don’t automatically die.
Ali’s case for Glean (“both app and a platform”): organizations are an n-squared coordination-overhead problem — “docs and Excel sheets and PowerPoints and meetings is how we move companies forward” — and much of that overhead is automatable.

6. Software isn’t dead — data entry is the wedge, and Zoom is the dark horse

Arvind rejects Satya’s crud-apps framing as “oversimplification”: Salesforce is “a full ecosystem of workflows,” and dynamic AI-generated UIs on a database won’t displace it because “most times you actually won’t know what you want” — good software companies design the interaction.
Ali’s sharper wedge: the big thing is data entry — how data appears in the system of record. “A company that would be well positioned would actually be Zoom… that’s where you’re having all the conversations. If you had that, that would be the full disruption of the SaaS.” Arvind confirms it’s already one of Glean’s most common agents: meeting recording → action items → Salesforce notes updated.
The episode’s best image of sprawl: the host joined a meeting with “four humans and six AI note takers”; the host heard of 17 in one discussion. “It felt like the first scene of a movie where the AI takes over.”

7. Inside the CEOs’ own stacks — change management is the bottleneck

Databricks internally: “Raffi,” an agent that surfaces the right customer story on demand; heavy agent automation across a 6,000-person go-to-market org and 3–4,000-person R&D org; and finance moved “from Excel to Python largely” — but only after an external data-science team built the models, “because they had their Excel models and they’re very proud of them.” HR may be less far along.
Arvind’s favorite is his daily prep agent, but the deeper story is a changed instinct: as CEO he used to ask a question and “30 people would be put on the task” — now he asks Glean first. Ali’s version: “usually Glean nails it; if not, then I’ll spin up a 30-person team to have three meetings.”
Arvind’s meta-lesson: “you have to have that belief that AI is a good collaborator… even if you don’t save time for the first few months, you’re actually going to improve the quality of your output.”

8. Rapid fire: yes there’s a bubble — long speech and agents, short the coding hype

OpenAI and Anthropic both up in 12 months (Ali: “revenue will be up — I don’t really understand how stocks work”): ChatGPT and Gemini “on fire,” and coding “we’ve only eaten into a small portion of that market.”
Bubble: yes, mapped to the camps — “There is a superintelligence quest camp — I would be very worried there. The researchers are super sober and nobody cares about them… probably the ones that are right, unfortunately.” Camp three “is not spending huge amounts of capital.” The hard evidence: “startups with zero revenue worth 10, 20, 30 billion. That’s a bubble.”
Longs: Ali on agents and speech — “as long as you’re using a keyboard, we haven’t nailed speech,” but “we’re this close to completely eliminating keyboards.” Shorts, carefully hedged: Arvind calls coding “a little bit over hyped — I don’t know if I would short it, it’s still the future,” and customer-service automation a little bit over hyped. Brad’s long: proactive AI products that “bring the AI to them” — the shift from “5% power users to 100%.”
Glean’s endgame, off a $200M revenue run rate and $10M deals: a “very personal companion for every person in every company,” fully privileged and confidential, that knows your day, goals and ambitions and “works on tasks before you ask it to. Today you come to Glean; in the future, Glean comes to you.”

PhD Bodybuilder Predicts The Future of AI (97% Certain) [Dr. Mike Israetel]What are we scaling?