AGI was already here last year—companies lag because they won't record meetings
Ali Ghodsi argues that by his AMP lab's original AGI definition from 2009–2010, the standard was already met last year—but enterprises are stuck on the first step: a bank CEO flatly refused to let AI record meetings, and without context, even the strongest models can't help.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AGI already hit the target; everyone just quietly moved the goalposts
According to Ali Ghodsi, by the consensus AGI definition held at UC Berkeley's AMP lab in 2009–2010, that standard was already met last year. He went back to confirm with lab colleagues that he wasn't misremembering; they confirmed he wasn't—the industry just quietly pushed the goalposts back. He believes that if we stopped all AI progress now, current capabilities would already be enough to transform the entire world—but this means all intellectual work humans can do, not superintelligence as Bostrom defines it.
— Ali GhodsiFour specific problems block enterprise AI, not technology
Meeting with customers daily, Ghodsi found that companies mostly treat AI as an advanced Google Search and a code-writing tool—and even then get only 20–30% speedup on code. The real bottlenecks are four things, in sequence: First, no context—a major bank CEO explicitly told him ‘it's impossible to record meetings, never in this lifetime’, and without being in the room, AI is useless. Second, even if they get the data, they fear being hacked; opening access permissions brings insider threat and cybersecurity risk. Third, token costs rise exponentially, but no company's revenue scales exponentially. Fourth, using Anthropic's or OpenAI's platforms risks having data and IP absorbed into their next-generation models.
— Ali GhodsiAI data disproved the CFO's assumption about employee satisfaction
Databricks analyzed an employee survey with over ten thousand participants using its internal product Genie and found that ‘does the company listen to employee feedback’ had only 60% satisfaction. The CFO guessed on the spot that poor ‘work-life blend’ was dragging down this metric. Ghodsi immediately queried Genie; the system pulled correlation data proving the two were not causally related and even surfaced a specific manager: work-life blend scoring as low as 44, yet ‘company listens’ scoring as high as 77%—showing he pushed employees hard but genuinely responded to their concerns. Analysis like this would take a manual data team a week.
— Ali GhodsiOnly a handful of engineers use AI code factories
Databricks plotted the weekly code submissions of 2000 engineers and found a stark L-shaped distribution: a small group at the top had output far exceeding everyone else. Investigation showed these people were running automated code factories; the rest were not. More counterintuitively, when adding cost dimension, these top performers with high output weren't spending more tokens than others—because they could route simple problems to cheaper models and use open-source models to save money.
— Ali GhodsiStaying private protects team morale from stock swings
Ghodsi says the benefits of going public are easy fundraising and marketing exposure, but his bigger concern is emotional stock swings: after employees tie compensation to stock, a price drop makes them feel ‘you lied to me’ and they quit, then three months later results prove the judgment wrong and the price rebounds—the company gets stuck in repeated roller coasters. Now he provides liquidity by organizing annual internal tenders, and over recent years found the allocated tenders were never fully used, suggesting most employees aren't eager to cash out.
— Ali GhodsiTrue RSI requires four conditions; none are met yet
AI writing 98% of code doesn't mean recursive self-improvement (RSI) is happening, Ghodsi says—whether we can automate the last 2% doesn't matter. True RSI requires simultaneously meeting four conditions: next-generation models must train superlinearly faster; resource usage must drop in sync, not just stack hardware for linear gains; accuracy must improve universally across all tasks, not at single points; and these three must loop continuously, bringing training costs down from hundreds of millions to single dollars. He says model training is actually getting longer and requires more manual data refinement and reinforcement learning gym work—this flywheel simply doesn't exist.
— Ali GhodsiSuperintelligence needs robot hands; cyberattacks are the real current threat
Responding to news of a French supermarket's refrigerator being hacked causing milk to spoil, Ghodsi says this isn't superintelligence, just a standard security flaw with password set to admin/admin. He believes that even if all four RSI conditions were met, to truly threaten people it would need a killer robot actually breaking into the room and pinning someone down—physical capability like that doesn't exist now. The only real current risk is cyberattack.
— Ali GhodsiExtinction risk is near zero; cybersecurity risk is real
Many AI researchers know research but haven't seen much economic data, Ghodsi says, and get swayed by science fiction narratives—hence claims like ‘10% human extinction probability’ becoming cited as authority. He personally rates that risk near zero, but cybersecurity risk is real, and that's what Databricks is building Lakewatch to monitor. He also believes AI truly benefiting enterprises and driving GDP will take decades, not the months or couple years labs claim—because many large institutions haven't even crossed the first step of ‘recording meetings’.
— Ali GhodsiIn their own words · checked verbatim
I actually think we already have artificial general intelligence. I've been saying it since last year.
Ali Ghodsi9:07
They're just using it as a glorified Google search.
Ali Ghodsi12:18
So if the token cost is going up exponentially, is your revenue at your company going up exponentially?
Ali Ghodsi15:27
Just because 98% of the code is being written by AI, we don't have RSI. If it automates the last 2%, who cares?
Ali Ghodsi46:17
you would have to have a killer robot come into this room right now and wrestle me down
Ali Ghodsi48:18
that risk is you know close to zero i think the cyber risk is real and really bad things can happen without
Ali Ghodsi59:26
Figures
| Databricks annualized revenue | $7 billion | 2:02 |
| Databricks year-over-year growth | 80% | 2:02 |
| Databricks 2015 annual revenue | $1.5 million | 2:02 |
| Databricks engineers (output analysis sample) | 2000 | 29:51 |
Glossary
- ontology
- Knowledge graph of organizational relationships between people, meetings, and documents, built by Databricks Genie.
- RSI
- Hypothetical recursive self-improvement: AI improving itself to create exponential capability acceleration.
- paperclip maximizer
- Bostrom's thought experiment showing how misaligned superintelligence pursuing narrow goals could cause catastrophic outcomes.
How to listen
Founders and investors building enterprise AI, tracking Databricks' business model, and wanting to understand how frontline figures evaluate AI doom theories.
38:25-44:02 — the discussion of whether to go public mostly follows standard venture financing logic and contains relatively low information density.