I've been living through a "backbone to pyramid" journey the last year as my "drink from the Slack firehose" strategy stopped scaling, and would love for the answer to instead be "All you need are LLMs to scalably process the context". I fear that we are underestimating the level of transformation embedded in the translation. Yes, some translations are mechanical and merely a source of friction to be eliminated. OTOH, my experience has been that most innovation emerges from the minds of human translators who are able to successfully span two domains and connect dots in a way that no one contemplated before. Are LLMs -- properly harnessed -- up to that challenge? Does the workforce have the human capital to make optimal use of the LLM output? I guess we're going to find out. It may well be that the companies which master the right context engineering for this problem will gain enormous edges in human capital efficiency.
Transformers — models that turn one sequence of symbols into another — came from the practical problem of translation at Google Translate. That was the first real-world motivation. At the time, no one thought of translation as a stand-in for intelligence. Searle’s Chinese Room gave us a hint, but I don’t recall anyone making the connection. Now it’s clearer: LLMs were built to translate. Their intelligence was an accident. Intelligence being “just” translation is the big surprise.
The category theorist would call these "structure-preserving transformations" and of course the tricky bit is preserving the right structures and not worrying about the extraneous/noisy ones: what information is signal and what is irrelevant?
Indeed that is the question. Is there really such a thing as "structure-preserving transformations"? Even the mythical Babel Fish would not preserve everything.
There could be many many "so what memos" that can be "translated" from a 20 page research paper. The "structure" that is worth preserving could be different in all of them.
This is a truly profound post. It gets to the heart of long standing inefficiencies in the enterprise knowledge work place. We thought the battle was won by simply digitizing paper processes and moving to tools like email, spreadsheets and slide decks. But that didn't eliminate the large amount of human translation work needed to handle enterprise processes. As you describe, LLM's have the ability to be transformational by dramatically reducing this translation work. However, to do so, requires a fundamental rethink of processes, underlying data and how LLM's work most effectively. An Everything As Code (EaC) mentality is needed to make our digital artifacts "machine actionable". This leverages standards like Json and Yaml. Many folks outside the tech space will have a hard time wrapping their minds around these ideas.
I love the thought process here, especially your line: "Creation and execution are why a business exists". Having been an entrepreneur and run my 2 startups and been involved with many in the past, I see a huge gap between creation and execution, and LLMs have certainly been filling it. In the future, the gap will be leaner. However, multiple roles between creation and execution will still be required because we live in a human world. With humans perfection isn't the key. In fact, being imperfect is what makes us different and better, and for this very reason we'll always need some translators :)
"It’s TRANSLATION: taking something in one form and turning it into another form that someone else can act on.
A soap formula becomes a manufacturable soap bar, then a marketing claim, then a retail SKU, then a line on a P&L."
I wonder if "Translation" applies to the other examples you cite, e.g., long doc into summary, but not so much the sequence above, especially if "manufacturable" means "manufactured." It seems to me that TRANSFORMATION is more apt. If your thesis were expanded from LLM to LLM + robotics (driven by AI), then there might be complete transformation of customer requirements into their satisfaction + profit. The day of Transformation in our economy is probably not immanent but likely not far out in the future either. I'm completely on board that the nature of all aspects of work are going to change (actually already in the process of changing). Thank you for sharing your ideas.
I think the organizations will still be shaped like a pyramid, but with different dimensions. You can't get rid of middle management in one fell swoop. But a million person organization where the average manager has 4 people reporting to them is much more "middle-heavy" than a million person organization where the average manager has 20 people reporting to them.
The question then becomes, do LLM tools make a middle manager's job easier. Can they manage 20 people instead of just 4?
Instead of collapse, might the amount of translation work explode as more translators join the mix? Won't there be more broken handoffs to fix? Won't some agents add to the ranks of the "leaders" and "doers" in your diagram, further increasing the "translation" load?
My guess would be that the total number of organizations (each with their own leader, doers) will proliferate because LLMs enable more middle management "translation" resources (both human and machine) to be shared by more organizations.
In your picture, that might look like the left-hand-side translators doing more translation but as fractional translators working for many organizations at once. Organizations will vary in their degree of structure as always, but the real change is not within a given org so much as across the space of all orgs.
It’s true: the cost of translating has gone down drastically - but solutions building the same for language (e.g. Palantir) are even better (also for LLMs) - LLMs still need to figure out exactly what was meant and there’s room for faults, friction etc.
I've been living through a "backbone to pyramid" journey the last year as my "drink from the Slack firehose" strategy stopped scaling, and would love for the answer to instead be "All you need are LLMs to scalably process the context". I fear that we are underestimating the level of transformation embedded in the translation. Yes, some translations are mechanical and merely a source of friction to be eliminated. OTOH, my experience has been that most innovation emerges from the minds of human translators who are able to successfully span two domains and connect dots in a way that no one contemplated before. Are LLMs -- properly harnessed -- up to that challenge? Does the workforce have the human capital to make optimal use of the LLM output? I guess we're going to find out. It may well be that the companies which master the right context engineering for this problem will gain enormous edges in human capital efficiency.
Transformers — models that turn one sequence of symbols into another — came from the practical problem of translation at Google Translate. That was the first real-world motivation. At the time, no one thought of translation as a stand-in for intelligence. Searle’s Chinese Room gave us a hint, but I don’t recall anyone making the connection. Now it’s clearer: LLMs were built to translate. Their intelligence was an accident. Intelligence being “just” translation is the big surprise.
The category theorist would call these "structure-preserving transformations" and of course the tricky bit is preserving the right structures and not worrying about the extraneous/noisy ones: what information is signal and what is irrelevant?
Indeed that is the question. Is there really such a thing as "structure-preserving transformations"? Even the mythical Babel Fish would not preserve everything.
There could be many many "so what memos" that can be "translated" from a 20 page research paper. The "structure" that is worth preserving could be different in all of them.
Excellent writing. I am indeed of those people trying to flatten that pyramid.
Insightful piece of work; contributes to the conversation.
Misses the integrative aspect of some kinds of work, characterized by the span of control concept encompassed in the USMC’s “Magic Number.”
(The Magic Number is 3+1)
This is a truly profound post. It gets to the heart of long standing inefficiencies in the enterprise knowledge work place. We thought the battle was won by simply digitizing paper processes and moving to tools like email, spreadsheets and slide decks. But that didn't eliminate the large amount of human translation work needed to handle enterprise processes. As you describe, LLM's have the ability to be transformational by dramatically reducing this translation work. However, to do so, requires a fundamental rethink of processes, underlying data and how LLM's work most effectively. An Everything As Code (EaC) mentality is needed to make our digital artifacts "machine actionable". This leverages standards like Json and Yaml. Many folks outside the tech space will have a hard time wrapping their minds around these ideas.
Worse than "not eliminating the large amount of human translation work needed", didn't those tools add to it?
We see AI everywhere but in the productivity statistics.
I love the thought process here, especially your line: "Creation and execution are why a business exists". Having been an entrepreneur and run my 2 startups and been involved with many in the past, I see a huge gap between creation and execution, and LLMs have certainly been filling it. In the future, the gap will be leaner. However, multiple roles between creation and execution will still be required because we live in a human world. With humans perfection isn't the key. In fact, being imperfect is what makes us different and better, and for this very reason we'll always need some translators :)
"It’s TRANSLATION: taking something in one form and turning it into another form that someone else can act on.
A soap formula becomes a manufacturable soap bar, then a marketing claim, then a retail SKU, then a line on a P&L."
I wonder if "Translation" applies to the other examples you cite, e.g., long doc into summary, but not so much the sequence above, especially if "manufacturable" means "manufactured." It seems to me that TRANSFORMATION is more apt. If your thesis were expanded from LLM to LLM + robotics (driven by AI), then there might be complete transformation of customer requirements into their satisfaction + profit. The day of Transformation in our economy is probably not immanent but likely not far out in the future either. I'm completely on board that the nature of all aspects of work are going to change (actually already in the process of changing). Thank you for sharing your ideas.
I think the organizations will still be shaped like a pyramid, but with different dimensions. You can't get rid of middle management in one fell swoop. But a million person organization where the average manager has 4 people reporting to them is much more "middle-heavy" than a million person organization where the average manager has 20 people reporting to them.
The question then becomes, do LLM tools make a middle manager's job easier. Can they manage 20 people instead of just 4?
True.
It's the "Taste" which diffentiate us, always :)
Instead of collapse, might the amount of translation work explode as more translators join the mix? Won't there be more broken handoffs to fix? Won't some agents add to the ranks of the "leaders" and "doers" in your diagram, further increasing the "translation" load?
My guess would be that the total number of organizations (each with their own leader, doers) will proliferate because LLMs enable more middle management "translation" resources (both human and machine) to be shared by more organizations.
In your picture, that might look like the left-hand-side translators doing more translation but as fractional translators working for many organizations at once. Organizations will vary in their degree of structure as always, but the real change is not within a given org so much as across the space of all orgs.
It’s true: the cost of translating has gone down drastically - but solutions building the same for language (e.g. Palantir) are even better (also for LLMs) - LLMs still need to figure out exactly what was meant and there’s room for faults, friction etc.