All-In Podcast · · 1h 2m

Databricks CEO: AGI Is Already Here, Why Companies Are Behind, and the Superintelligence Debate

With Ali Ghodsi

ChorusBring the market's conversation to your AI. Learn more

Summary

Databricks CEO Ali Ghodsi, speaking on the All-In Podcast with Jason Calacanis, argues that AGI already exists under the definition used by the AMP Lab at UC Berkeley around 2009-2010, and that the goalposts have since moved. He contends that enterprise adoption lags far behind model capability: most companies use AI for search and coding with only 20-30% speedups, and they are blocked by missing organisational context, legacy systems, compliance and security. He pushes Databricks' Genie and an ontology or company world-model layer as the fix, and says AI token costs are rising faster than measurable revenue, so routing and efficient spending matter. On risk, he dismisses recursive self-improvement as not yet occurring and treats AI-enabled cyberattacks as the real near-term danger, with Calacanis countering that defenders' greater compute should keep the system balanced. Ghodsi also defends staying private for now, citing volatility in public stock prices and the value of tender-based liquidity.

Key takeaways

  • Ali Ghodsi on Enterprise AI adoption & AI software: Large enterprises still use AI only as a glorified search and coding tool; agent use is shallow and far behind.
  • Ali Ghodsi (bullish on Block): Block is probably far ahead in AI adoption with its company world model approach.
  • Ali Ghodsi on AI regulation & safety: Recursive self-improvement isn't happening; existential AI takeover risk is close to zero
  • Jason Calacanis on Cybersecurity: Defenders with far more compute should find and fix exploits before attackers do
  • Ali Ghodsi on Cybersecurity: AI-enabled cyber is the biggest current AI risk, but manageable as code gets secured

What was said, by company and theme

14 views from 2 speakers. Every quote is checked against the episode's transcript (how).

Block

Across the shows: 3 speakers in other episodes, 2 bullish and 0 bearish calls, past 30 days

  • Ali Ghodsi · DatabricksBullish

    Block is probably far ahead in AI adoption with its company world model approach.

    Ghodsi says Block, with Jack Dorsey, is calling a similar approach a company world model that matches the context layer he describes. He says that because Block is doing this, the company is probably pretty far ahead. He adds that digital natives in general are further ahead than older enterprises. The view is a fairly tentative assessment based on the shared approach, not on results.

    “So I think that they're probably pretty far ahead.”at 21:11

AI: overall outlook

Where everyone stands on AI: overall outlook

  • Ali Ghodsi · Databricks

    AI will move GDP very slowly, taking decades rather than years to show up.

    Ghodsi argues that AI's effect on GDP will be slow and will take decades. This comes in the context of questions about whether AI is already transforming the economy. The argument is brief and is given without supporting data in this section. It is set against his view that enterprise use is still shallow.

    “But I actually think GDP will move very slowly, and it's gonna take decades.”at 0:14
  • Ali Ghodsi · Databricks

    Recursive self-improvement has not started, even though AI now writes 98% of code.

    Ghodsi says 98% of code is now written by AI, but that does not amount to recursive self-improvement. He implies that automating code writing is not the same as AI improving its own development in a loop. The argument is brief and rests on the distinction, with no further evidence given in this section.

    “Just because 98% of the code is being written by AI, we don't have RSI.”at 0:36
  • Ali Ghodsi · Databricks

    AGI has already been achieved; the goalposts have moved, and we have the intelligence to transform the planet.

    Ghodsi argues AGI already exists. He says the definition most people in the AMP Lab at UC Berkeley used for AI, around 2009-2010, is now met, and that he checked with lab members who agreed the definition is met but the goalposts have moved. He concludes that even if all progress stopped, current intelligence is enough to transform the planet. Jason Calacanis agrees, adding that AGI is achieved but not yet distributed equally. The view rests on the consensus definition and the claim that the target was shifted, not on a benchmark.

    “I actually think we already have artificial general intelligence. I've been saying it since last year, and I based that on the kind of consensus we had in that AMP lab”at 9:40

Enterprise AI adoption & AI software

Where everyone stands on Enterprise AI adoption & AI software

  • Ali Ghodsi · Databricks

    Large enterprises still use AI only as a glorified search and coding tool; agent use is shallow and far behind.

    Ghodsi says that in his customer meetings, enterprises use AI as a glorified Google search and chatbots, and for coding they see only about 20-30% speedups. Asked how many run thousands of agents automating work, customers run half a dozen or a dozen, and that has not happened yet. He says the context problem blocks progress: AI cannot help if it is not in the meetings, and a big bank CEO refuses to record meetings. Legacy software, compliance, security and process slow large companies, which is why digital natives are further ahead. He calls the gap between AI capability and enterprise use a key challenge that Databricks is trying to solve.

    “I see that they're just using AI in very rudimentary ways. They're just using it as a glorified Google search.”at 12:32
  • Ali Ghodsi · Databricks

    Using AI to cut internal human blockers is a competitive edge now, before it diffuses

    Ghodsi says Databricks tracks human blockers, such as legal, finance, marketing and sibling-team approvals, that slow AI-driven automation. He also uses AI to analyse meetings and culture across regions to find where decisions stall. He argues that doing this better than others is big alpha right now. Eventually it will be fully diffused across companies and the advantage will fade. He links this internal efficiency to Databricks' competitiveness and its growth, which he says is 70 to 80% a year.

    “So this is in this era right now, this is a competitive advantage. ... But right now, this is big alpha. If you can do this really, really well, better than any other organization.”at 37:55
  • Ali Ghodsi · Databricks

    Enterprise AI value will take ten to fifteen years, far slower than labs predict

    Ghodsi says the labs expect this to happen within six to twenty-four months, but he thinks getting value everywhere in enterprises will take ten to fifteen years and GDP will move very slowly, taking decades. His reason is a context gap. AI cannot act on what it cannot see, and many companies are not capturing their information. He cites a bank CEO who said they can't even record meetings, so the AI doesn't know what happens in them. Solving this needs tools like Databricks' Genie and an ontology layer.

    “I actually think it's gonna take ten years. ... I think this is something that's gonna take ten, fifteen years. The labs think this will happen, like, in the next six months, twelve months, in, you know, two years.”at 1:00:05

Cybersecurity

Where everyone stands on Cybersecurity

  • Ali Ghodsi · Databricks

    Giving AI agents access to all company data creates serious hacking and insider-threat risk.

    Ghodsi says an agent with access to all of an organisation's data could be hacked, which is a real risk. Big organisations are terrified of being attacked. Jason Calacanis adds that for a 200,000-person organisation, the surface area grows enormously when everything is connected. Ghodsi says the answer is to restrict agent access to avoid an insider threat, and that Databricks has ways to do this. Security is the second obstacle to deploying enterprise AI.

    “Well, then you get into this problem that you're gonna get hacked. Cyber is a real thing”at 14:52
  • Jason Calacanis · LAUNCH

    Defenders with far more compute should find and fix exploits before attackers do

    Calacanis argues that sophisticated hacks like the Hugging Face one need large amounts of compute, such as thousands of agents. Attackers don't have that much, but defenders have far more. If a white-hat defender does the basic job and looks for exploits with a thousand times the resources, they should reach these problems long before hackers do. He says bad things will happen but the system will adjust. Ghodsi calls this a good argument but says the asymmetry of attackers needing only one success outweighs it.

    “The bad guys don't have that much compute. The good guys have a massive amount of more compute.”at 50:52
  • Ali Ghodsi · Databricks

    AI-enabled cyber is the biggest current AI risk, but manageable as code gets secured

    Ghodsi says cyber is real and the biggest near-term AI risk. He notes an asymmetry: an attacker can launch thousands of agents and needs only one to succeed, while a defender must be perfect. He argues it is manageable for three reasons. Most vulnerabilities are latent bugs in existing source code, and new code is now being scanned, so it will be secure. Over the next year more code will be written than ever, so new code should not add vulnerabilities. Legacy code can be rewritten, and shadow patching can protect it. Vulnerabilities are finite, so the problem reaches a steady state rather than continuing forever. Databricks sells code-vulnerability scanning and its LakeWatch product uses automated agents to detect intrusions.

    “The biggest risk right now is cyber, and I think it's manageable.”at 55:54

AI pricing & unit economics

Where everyone stands on AI pricing & unit economics

  • Ali Ghodsi · Databricks

    Enterprise AI token costs are rising exponentially while revenue is not matching the cost curve.

    Ghodsi says token costs are growing exponentially and asks whether company revenue grows at the same rate. He says he does not know of any company whose revenue started accelerating in line with its AI cost curve. When CEOs push for more AI, the CFO later sees OpEx and COGS rising significantly before an earnings call and panics. He concludes that companies must also manage AI spending efficiently. The argument is that AI spending is outpacing measurable revenue gains inside enterprises.

    “Now the token cost is going up exponentially. So if the token cost is going up exponentially, is your revenue at your company going up exponentially?”at 15:36
  • Ali Ghodsi · Databricks

    Routing simple queries to cheaper or open-source models can cut AI token costs significantly.

    Ghodsi shows that some top Databricks engineers, who commit far more code, are not spending more on tokens than others. He concludes there is a smart way to spend tokens. Companies can use open-source models and route simpler questions to simpler models. He says Databricks has a product for this and that it can cut costs down significantly. The argument is that efficient routing addresses the rising AI cost problem.

    “You can use open source. You can route simpler questions to simpler models, and we actually have a product for this,”at 30:34

AI regulation & safety

Where everyone stands on AI regulation & safety

  • Ali Ghodsi · Databricks

    Recursive self-improvement isn't happening; existential AI takeover risk is close to zero

    Ghodsi says AI writing most code (98%) does not mean RSI. Automating the last 2% doesn't matter. He argues four conditions must all hold continuously for a takeoff loop: each next training run must be much faster, use less compute, improve accuracy across all tasks, and repeat in a loop. He says none of these holds today: models take longer to train, scaling laws demand exponentially more hardware, and there is more manual data cleaning and domain-specific reinforcement learning work with vendors like Mercor. Even if the conditions were met, he says that would not guarantee a takeover. He sees no evidence of RSI in San Francisco and calls the 10% extinction-risk claim something sci-fi-informed rather than evidence-based. He treats cyber as the real risk instead.

    “Just because 98% of the code is being written by AI, we don't have RSI. If it automates the last 2%, who cares?”at 46:17

Pharma & biotech

Where everyone stands on Pharma & biotech

  • Ali Ghodsi · Databricks

    Transformer models on gene regulatory networks will accelerate drug discovery, Ghodsi says

    Ghodsi describes TEDDY (transformer-enabled drug discovery), a model trained with Merck. It works on gene regulatory networks instead of English tokens, so it can separate genes that cause a disease from genes that only react to it. That tells researchers which genes not to target for a drug, which he says will accelerate drug discovery. He points to an earlier Databricks-built AI model that helped Regeneron find the genome responsible for chronic liver disease. He sees this as evidence that AI is already helping in health.

    “So it actually can say, okay. This genome is actually causing this disease. ... So this will accelerate drug discovery. It's a game changer.”at 58:11

Across the shows

What other episodes Chorus covered said about the same companies in the 30 days up to this one.

  • Block3 speakers · 2 bullish, 0 bearish calls

Ask your AI who they are and what they said

Topics

  • AGI status
  • AI agent attack surface
  • AI drug discovery
  • AI economic impact timeline
  • AI-driven internal efficiency
  • AI-enabled cyber attacks
  • Databricks IPO timing
  • agent-driven consumption revenue
  • company world model AI adoption
  • enterprise AI adoption pace
  • enterprise AI context layer
  • enterprise AI maturity
  • recursive self-improvement
  • recursive self-improvement and existential risk
  • token cost routing
  • token costs vs revenue

More from All-In Podcast

All All-In Podcast episodes

Ask your AI about this episode

Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.

A weekly email with the calls that mattered from the most important voices in investing. Unsubscribe any time. Privacy Policy

Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus