He equates deploying AI to hiring capable but clueless workers.
Garry Tan#121
Imagine replacing 90% of your employees with a team of geniuses who have no idea how your company operates.
Total chaos. Nothing works.
That’s what AI feels like today.
The missing piece is extracting all the domain knowledge from people’s heads and providing that as structured context to the models.
Many users agree that extracting company tribal knowledge is the top barrier to enterprise AI adoption because it enables effective agents, while some criticize the job-replacement framing as anti-human.
No Digg Deeper questions have been answered for this story yet.
This is the actual bottleneck. The models are smart enough already. What is missing is the company-specific context locked in senior people heads. Whoever cracks knowledge extraction at the company level unlocks the rest.
As you work on this, please consider using GBrain as your OSS retrieval layer
This is effectively the #1 problem for AI agents in the enterprise.
As we go from agentic coding (where a large amount of context is in the code base, and users are technical enough to get the rest to the agent easily) to a world of knowledge work agents, the context problem becomes much more acute.
We see this every day with customers at Box. For existing digital knowledge, it’s often fragmented across legacy systems or environments that don’t play nice with agents, and have access controls that don’t map to the real work that needs to be done, which become a huge hurdle for getting agents the context they need. This has to all get moved to modern, secure cloud environments.
But also, companies often haven’t captured and digitized some of the critical context that agents need to work with. Decisions, processes, and workflows often live in people’s heads and tribal knowledge that need to get turned into unstructured data for agents.
This is actually one of the biggest points of leverage for applied AI companies, because they can work to specialize in getting agents exactly the information and domain expertise they need. But it’s also one of the reasons why FDEs and new system integrator plays will also work so well right now.
The companies that figure this out will be able to get the most out of AI going forward.
Imagine replacing 90% of your employees with a team of geniuses who have no idea how your company operates.
Total chaos. Nothing works.
That’s what AI feels like today.
The missing piece is extracting all the domain knowledge from people’s heads and providing that as structured context to the models.
@t_blom This problem will naturally tend to go away as companies are grown from the start using AI. Then you don't need to extract any domain knowledge from people's heads; it will never have been in people's heads.
Imagine replacing 90% of your employees with a team of geniuses who have no idea how your company operates.
Total chaos. Nothing works.
That’s what AI feels like today.
The missing piece is extracting all the domain knowledge from people’s heads and providing that as structured context to the models.
Agreed!
This is what Software Factory helps enable - it organizes and creates the artifacts that guide all downstream productivity.
Increasingly we are finding that large enterprises want a control plane that organizes the tribal knowledge while also providing governance, safety and control.
This seems to be accelerating as the models asymptote on quality.
Try it here: http://8090.ai
This is the actual bottleneck. The models are smart enough already. What is missing is the company-specific context locked in senior people heads. Whoever cracks knowledge extraction at the company level unlocks the rest.
As you work on this, please consider using GBrain as your OSS retrieval layer
@t_blom It would be interesting, and a good opportunity for startups, if this requires new tools to be made. But no one will be able to figure out faster what these tools should look like than the startups that need them.
@garrytan @garrytan we agree, we built HQ. It’s the shared company context layer built on top of Claude Code, Codex, Openclaw, Hermes’ etc
http://getindigo.ai
Hey Garry, GBrain already covers more than half of the lift for this already!
There are 2 main components: 1. Memory indexing (embedding/enrichment) during ingestion 2. Accurate, Low latency, cost-effective Retrieval using the enriched embeddings
I measured GBrain across industry benchmarks LoCoMo, LongMemEval, and the “hardest” one Memory Agent Bench (MAB) — did you know your system crushes some of the leading Labs out there on MAB?
Happy to share the data any time. Invested like $500 to run your system on those benchmarks and get the data because I was genuinely curious. Nice work ser!
DM if you’d like and I’ll send the benchmark data right on over.
@garrytan I'm seeing a big bottleneck for multiplayer UX, think this will be a huge opportunity
@garrytan I have a theory on some of the constraints. And a partial solution
@garrytan Senior people sit down, talk with the LLMs for a few hours using voice. Random prompts given from the knowledge you want extracted, annoying initially but solves this.
@garrytan @garrytan completely agree
we built @hyperspell as the company brain to extract tacit knowledge from granola, slack, email, and structure it as a filesystem for agents
our retrieval layer uses some of the concepts from gbrain and it’s a game-changer
http://hyperspell.com
The modern corporation was always limited by the "headache" of retaining talent who owned the critical domain expertise. AI natively-grown startups fix this interface bug. If the core intelligence is institutionalized inside the LLM repository from Day One, the "talent war" is officially over. Capital wins by default.
Yeap, while writing/reading scientific papers you realise that lots of know-how, failed experiments and domain expertise are never shared. The creative process is locked within tacit knowledge. I wanna learn about the hidden motives, the biases, the reasoning, the aha moments, etc. Whoever unlocks this, unlocks the next wave of artificial creativity. I was thinking on interviewing scientists to extract and collect this data. I guess applies to senior heads as well.
@garrytan "Models are smart enough" is the comfortable read. The real bottleneck is we make them prove it one prompt at a time instead of trusting them with a whole job. The gap is permission, not intelligence.
@garrytan we believe the jump happens in capturing it while the agent works and not as a separate "harvesting" project. every tool call, every decision, the agent writes what happened in the loop. extraction must stop being a project and become a byproduct of the work.
It's a transitional problem.
Possibly, a better approach could be to have an agent that shadows a senior employee with LLM judge that processes the context and churns out pattern analysis reviewed by a human-in-the-loop if that was the basis of the decision.
Eventually, building a knowledge-base version for that department with the other subagent monitoring team, and then merging in the department-level knowledge.
@garrytan this will help facilitate the knowledge graph
@paulg @t_blom even if the company is all agents from day one, you still need a company brain to centralize and share context between agents
the source is different but the need remains
Exactly - Knowledge extraction is the bottleneck.
But once that knowledge enters agentic systems, the next bottleneck appears: trajectory integrity.
A system can retrieve the right context and still drift in how it applies it over time.
Enterprise AI will need more than memory and retrieval.
It will need external observability of intent, context, constraints and runtime behavior.
That is where ASA + ASC fit:
ASA observes drift. ASC stabilizes bounded response paths.
Knowledge retrieval unlocks the company. Trajectory stability keeps it safe.
Mieczysław Kusowski HumanAI / Symbioza2025
@garrytan Seems best for self-hosted option
