Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 24 August 2026.
At this point an AI post on Hacker News barely needs a subject. The subject is a match dropped into a warehouse full of prewritten arguments.
Imagine the submission:
Claude can now edit Excel files
487 points | 296 comments
The first comment is useful.
tableflip42
I tried this on a twelve-tab operating model. It fixed some ugly formatting, added the formulas I asked for, cleaned up a few labels, and found two cells where somebody had hard-coded a number into a formula column. Pretty impressive. I would still review every change, but this is already useful for tedious spreadsheet work.
For about four minutes, we are discussing Excel.
oldunixguy
I gave it a workbook with named ranges and external references. It broke a couple formulas, invented a sheet name, and confidently told me the workbook was consistent. Took longer to inspect the damage than to do the work myself.
tableflip42
Sure. I would use this for bounded edits with review, same as coding agents.
oldunixguy
Then the demo should say "bounded edits with review" instead of implying it can operate spreadsheets.
bayesian_burrito
Both experiences can be true. The interesting question is the distribution of task difficulty and failure cost.
Ah, okay. We're off.
Somebody now has to say that humans make mistakes too.
latent_space_cadet
People keep treating human spreadsheet editing as the reliable baseline. Finance departments are full of broken formulas made by people. The relevant comparison is error rate per unit of useful work.
oldunixguy
A human analyst who breaks a formula has a model of why the formula exists.
latent_space_cadet
Have you met humans?
This gets 146 points.
Somebody else shows up to explain that the real issue is tool use.
promptartisan
I think a lot of the bad experiences come from asking the model to "fix the spreadsheet" instead of giving it a precise task and a verification loop that checks the result. These systems reward good decomposition.
cobolsaint
If using the labor-saving device requires me to become an expert supervisor of the labor-saving device, some of the labor has moved. The disappearance was oversold.
promptartisan
Learning to use a compiler also required learning a new tool.
Oh, great. The compiler is here.
Of course it is.
deterministic_ferret
A compiler does the same thing given the same source and toolchain. That analogy gets weaker every time people use it.
promptartisan
Git, databases, networks, browsers, package managers, and humans all have failure modes. Professional work has always involved managing imperfect abstractions.
deterministic_ferret
We are now calling a coworker who invents APIs an abstraction layer.
Somewhere around here the spreadsheet becomes a referendum on the philosophy of tools.
Then somebody posts a benchmark.
evalmaxxer
The anecdote war is getting old. On SpreadsheetBench 2, the new model is up 18 points over last year's best system. That is a huge move in twelve months.
contamination_station
Benchmark is probably contaminated.
evalmaxxer
Evidence?
contamination_station
The tasks are public.
evalmaxxer
Public tasks can still measure capability if the held-out variations are designed correctly.
gradient_grump
Benchmarks are becoming product marketing with error bars.
evalmaxxer
And vibes from one failed workbook are science?
At this exact moment, somebody links an arXiv paper. Somebody else has already read the appendix. Then someone finds a sentence on page 47 saying the model had access to LibreOffice during evaluation, and the replies split over whether this ruins the result or makes it more realistic.
Then somebody says the one fucking word that can ruin the rest of your afternoon:
scale_pilled
The larger point is the trajectory. A year ago these models could barely do this at all.
Oh, for fuck's sake. Somebody said trajectory.
Great. We're never seeing the spreadsheet again.
diminishing_returns
People have been extrapolating curves through saturation points for four years.
scale_pilled
And the systems keep getting better.
diminishing_returns
Better at benchmarkable tasks.
scale_pilled
Which keeps including more tasks.
recursive_skeptic
Every time a model gains a capability, that capability gets reclassified as irrelevant to intelligence.
syntaxerror69
Every time a model fails a capability, that failure becomes an engineering problem scheduled to disappear next release.
Both comments receive hundreds of points. Everybody feels seen.
A little further down, Excel becomes a proxy war for software engineering.
senior_staff_plus
I use these tools all day. They are excellent for boilerplate, unfamiliar APIs, test generation, migrations, and getting a first pass over code I already understand. They save me hours every week.
vim_in_my_blood
I tried one for a refactor last month. It produced 900 lines of plausible sludge and missed the invariant that mattered. I spent more time reviewing it than I would have spent writing the code.
senior_staff_plus
That sounds like a task I would keep human-led.
vim_in_my_blood
So the senior engineer gets the easy repetitive work automated and keeps the hard reasoning. Great. How does the junior become senior?
Ah, fuck. Here come the juniors.
apprentice_problem
This is the part the industry is sleepwalking into. Senior engineers can use AI because they already know what good code looks like. Companies see the productivity gain and hire fewer juniors. Ten years later, where do the seniors come from?
founder_mode_on
Juniors with AI can take on larger tasks earlier. The apprenticeship changes.
apprentice_problem
Earlier access to larger tasks only helps if somebody teaches them how to tell good output from garbage.
founder_mode_on
That was already true of Stack Overflow.
greybeard_1978
I learned C from K&R, man pages, and a machine that crashed when I was wrong.
Nobody asked how he learned C. Everybody knew he would tell us.
Give it a little longer and now everybody has an economics degree.
comparative_advantage
Even if AI makes programmers more productive, demand for software could rise enough to absorb the gain. We have seen this with prior productivity tools.
typesetter_union
We have also seen occupations disappear.
comparative_advantage
Software demand is unusually elastic.
budget_owner
Corporate budgets are unusually finite.
jevons_paradox_enjoyer
If cost per unit of software falls, consumption can rise dramatically.
budget_owner
Jevons paradox still stops short of every company buying infinite software.
Someone says Baumol. Somebody else says induced demand. A person with "econ" in the username appears and spends twelve comments correcting everybody's terminology, then leaves behind twelve corrections and zero forecasts anyone can check later.
Then salaries arrive.
remote_nomad_88
The immediate effect may be global wage compression. If one senior engineer with agents can do the work of a larger team, companies need fewer people and have even more reason to hire globally.
sf_realist
Or the best engineers become dramatically more valuable because the leverage compounds with skill.
remote_nomad_88
Both can happen. Fewer people, higher pay for the top end.
solidarity_dot_exe
Amazing how every technology discussion eventually produces a theory explaining why the speaker personally deserves a larger share of the surplus.
That one gets 287 points and a small founder pile-on.
Remember Excel?
Nobody else does.
Eventually somebody says "stochastic parrot."
parrot_skeptic
The core limitation remains. The model predicts tokens. Token prediction gives me zero evidence that it understands the spreadsheet or the consequences of its edits.
predictive_brain
Human brains also predict sensory input.
parrot_skeptic
This argument is doing incredible work with the word "also."
predictive_brain
The point is that prediction and understanding are compatible.
embodied_symbols
Genuine semantics requires embodiment.
chess_engine_reply_guy
AlphaGo had no body.
embodied_symbols
Go has a complete formal environment.
chess_engine_reply_guy
Excel is beginning to look pretty formal from here.
Somebody mentions the Chinese Room. Somebody explains that everybody misunderstands the Chinese Room. Somebody else explains that the explanation misunderstands the Systems Reply. Another person says consciousness is irrelevant to economic usefulness. Then someone says economic usefulness without consciousness is exactly what makes the whole thing dangerous.
Ah, and there we are: AGI.
Took a while, actually. Hacker News is slipping.
pdoom_17pct
The spreadsheet demo itself is trivial. The concern is what happens if systems capable of increasingly general digital work continue improving. At some capability level, alignment and control dominate the discussion.
secular_eschatology
We have somehow gone from formula editing to machine gods in under 200 comments.
pdoom_17pct
Low-probability catastrophic risks deserve attention when the downside is enormous.
base_rate_bandit
Assigning 17% to an unprecedented event is numerology with decimals.
pdoom_17pct
Quantification makes the uncertainty explicit.
pascal_mugged_me
There is an entire literature on why tiny probabilities times giant utilities produce absurd decisions.
Now there are links to LessWrong, an old Nick Bostrom interview, a corrigibility paper, and some ancient forum post that somebody swears anticipated the entire debate. The replies immediately split over whether any of them belong in the same conversation.
Somewhere deep in the wreckage, the energy person arrives.
megawatt_monk
Everyone is discussing labor as if inference were free. The capex and power requirements for scaling these systems are enormous.
tokens_get_cheaper
Cost per token has fallen by orders of magnitude.
megawatt_monk
Total demand has risen even quicker.
nuclear_option
Good. Build nuclear.
Beautiful. We have now achieved Hacker News's final form: a nuclear-power thread hiding inside an AI thread hiding inside an Excel thread.
Then copyright arrives because copyright always arrives.
copyleftist
We are skipping over the fact that these models were trained on huge amounts of copyrighted work without consent.
transformative_use
Training is computational analysis of the work.
copyleftist
Tell that to the illustrator whose style gets reproduced on command.
latent_lawyer
Style itself is generally outside copyright protection. Specific outputs can still infringe.
artist_in_residence
I love when programmers discover one sentence of copyright law and become maritime attorneys of the imagination.
Somebody brings up books. Somebody brings up open source licenses. Somebody asks whether generated code can reproduce GPL fragments. Somebody pastes a seven-year-old court opinion and has misunderstood the holding. Actual attorneys eventually show up, disagree with each other, and leave.
Then the open-source model person says closed labs are the real problem. A few replies later somebody says open weights make catastrophic misuse easier. Soon somebody else says local models are the only defense against surveillance capitalism. Another person points out that running a 200-billion-parameter model locally is a peculiar definition of consumer freedom.
And way down there, after all of that, the person who actually used the spreadsheet feature comes back.
tableflip42
Update: I tried it on another workbook. It handled formula fills and formatting well, struggled with a pivot table, and gave up cleanly when I asked it to trace one external link. Pretty useful overall.
Two points.
Zero replies.
The metaphysics are elsewhere now.
Seven hours after the submission went up:
spreadsheet_dad
Does it preserve pivot table slicers?
Zero replies.
The thread has forgotten the Excel file.
The Excel file, somehow, remains the only thing anybody came here to ask about.