The IRS Is Buying an AI Supercomputer from Nvidia
theintercept.com
theintercept.com
Reading through the article, it's not one or two machines but 8 Blackwell nodes, so quite some firepower for compute. Wonder what part of the audit or tax prep requires a GPU farm.
They just sent you your assessment automatically with they you think you/they owe if anything, you need to file anything only there if you believe there is difference .
The US system is pretty backwards really, most of the taxes are deducted at source anyway the system already has a good idea of your income/taxes owed.
If IRS calculate returns at scale using tax deducted at source and employer info(W-2 etc) , unless you dispute their calculations there is no need for audit or paperwork at all reducing the cost of running IRS too.
This kills intuit so will never happen, but if it does it would be the single biggest revenue boost to the Government easily in tens if not 100's of billions of dollars.
Why should we have a party dragging out a transaction? Just complete it immediately and ideally in isolation.
Then they might as well also run the bank, the payment processor and the recruitment agency.
People here in NL are afraid to start a business not because they can't deliver the product or service or that they can't find customers, they are afraid of the tax office.
The IRS is the most bang for the buck government department. Every dollar spent there nets out many more, because they're chronically under-resourced to audit the whales who can trivially hide their tax obligations.
Naturally whenever there's a hint that the department could be even close to properly funded, the media, overwhelmingly owned by the class of people who evade taxes as a sport, gin up controversy. It's no wonder that the richest man in the history of the world has bought his way into the building and been provided with gasoline and matches.
Sounds like fun to find and fix them all again though!
[1] https://www.fastcompany.com/91278597/elon-musk-doge-cobol-la...
DOGE just shut down the auto prepare because gov’t is bad and wasteful.
This is just a trope.
Auditing big whales is often not the most efficient use of resources because their taxes are a) much more complicated, and so more expensive to audit and b) more likely to have been done by a team of competent lawyers, so even though they're engaged in all kinds of shenanigans, the shenanigans are legal and there is no money to be had by auditing them.
Where the IRS can "turn a profit" is by auditing small and medium businesses, because they're more likely to have screwed up, and auditing them is cheap to the IRS. But this has a different problem. It isn't cheap to the taxpayer being audited; they pay an order of magnitude more to go through the audit than the IRS does. And then most of them are innocent, even though the IRS makes it back from the few that aren't. For obvious reasons the innocent ones hate this with the fire of a thousand suns, because nobody compensates them for their costs even though they did nothing wrong. But if you actually did compensate them for their costs then it wouldn't be net-positive for the IRS to do all those audits.
Do you have citations to back up this assertion? The GAO seems to say otherwise[1].
> GAO found that, as income increased, IRS recommended more taxes per audit, on average, but generally closed fewer audits.
They close fewer audits because the audits require more resources (and partially because there are fewer of those taxpayers). Meanwhile the other effect also shows up clearly in their numbers: Average additional tax for the $500k-1M group (midpoint $750k) was ~$20k or 2.67%, for $1M-5M (midpoint 3M) was ~$40k or 1.33%, for $5M-10M (midpoint 7.5M) was ~$86k or 1.15%. Percentage for the top group is unknown because that bracket is open-ended, but $360k for the group that contains billionaires is likely to continue the trend.
They get less as a percentage from the people with better tax lawyers. It's more in absolute dollars because by definition they have more money, but since it also costs more to audit them it's not necessarily worth it.
Notice also that these are heuristics and the actual numbers change with the number of audits. There are going to be some people who are obviously cheating on their taxes, e.g. someone else filed a 1099 or W-2 for paying them and then they filed a return that didn't include the income. The IRS is going to put them at the front of the list to be audited and then that audit is going to show a positive recovery for the IRS.
The more audits you do, the lower the recovery rate gets, because you started with the ones with huge obvious problems and doing more audits means auditing people who are more likely to be innocent.
This is also why the "bang for the buck" calculation is pretty meaningless. Suppose there is one person you can audit for $2000 and then recover a million dollars (+$998,000), and then an unlimited number of people you could audit for $1000 and recover an average of $800 (-$200). If you do one audit of the one who is obviously cheating, you've recovered 500 times more than you put in! Clearly you should keep going. If you audit 100 people, you've spent $101,000 and recovered $1.08M, that's still almost ten times more than you spent! Clearly you should keep going? Except that only one of them was net positive and the additional audits have actually reduced your net gain from $998k to $978k while auditing a significant number of innocent people.
What matters is the incremental returns from additional audits, not the average returns from existing audits. And taking into account as a cost the burden of auditing innocent people.
For "x" hours of work, the big whales can also potentially bring in a lot more money than a bunch of small minnows.
Before it got a funding bump, the IRS went after minnows because of resource constraints:
> Congress asked the IRS to report on why it audits the poor more than the affluent. Its response is that it doesn’t have enough money and people to audit the wealthy properly. So it’s not going to.
* https://www.propublica.org/article/irs-sorry-but-its-just-ea...
Your article says the opposite:
> They [audits of low income people taking the EITC] are “the most efficient use of available IRS examination resources,” Rettig’s report says.
They can either hire a lot of cheap low-level examiners to quickly audit a lot of low-income people or a smaller number of expensive examiners to slowly audit big complex returns that are more likely to be correct but have more at stake. They were doing the first one because it recovers more money per enforcement dollar.
You can decide that your priority is "audit more rich people" rather than "recover the most money" but that doesn't make them the same thing, or make the first one the only or even default thing they'll do if you give them more enforcement resources.
Some of this is just a war against the middle class.
But more seriously this is a desire to let unregulated unchecked systems draw judgement upon people. This is like the UK postal crisis, where the UK has long held a legal system that assumed computerized systems are correct and not questionable.
Once the AI is in power, it can be tuned to the biases of the ultra-wealthy tech-right. Technocratic dominion.
The IRS is mentioned in Project 2025's chapter on Treasury:
* https://static.project2025.org/2025_MandateForLeadership_CHA...
See "TAX ADMINISTRATION" section. Some points with-in that:
> Doubling the IRS? The Inflation Reduction Act contains a radical $80 billion expansion of the IRS—enough to double the size of its workforce.34 Unless Congress reverses this policy, the IRS will become much more intrusive and impose still greater costs on the American people.
[…]
> Management. The IRS has approximately 81,000 employees.36 Of those, only two are presidential appointments—the Commissioner and the Chief Counsel.37 As a practical matter, it is impossible for these two officials to overcome bureau- cratic inertia and to implement policy changes that the IRS bureaucracy wants to impede. That is why, notwithstanding decades of sound and fury, almost nothing has changed at the IRS.
On the second point, "Management": who doesn't / wouldn't want more political apparatchiks with-in the tax system?
As for the second point: the IRS has a resource problem (partly due to retirements), and needs to hire a bunch of folks now to be able to train them and pass on institutional knowledge:
> “When we hear 87,000 agents, it sounds like a horrifying army of people. But it’s what they need to maintain, because people are exiting in volumes,” says Robert Cordasco, a certified public accountant and founder of Cordasco & Co. in Savannah, Georgia. “I don’t know how much it adds as much as it gets us to the status quo.”
* https://www.cnbc.com/2022/08/31/irs-is-not-hiring-an-army-of...
* https://www.journalofaccountancy.com/news/2024/jun/expert-wa...
A story from before the funding bump:
> Congress asked the IRS to report on why it audits the poor more than the affluent. Its response is that it doesn’t have enough money and people to audit the wealthy properly. So it’s not going to.
* https://www.propublica.org/article/irs-sorry-but-its-just-ea...
He literally gloated on his website about closing it
"IRS launches new effort aimed at high-income non-filers; 125,000 cases focused on high earners, including millionaires" [https://www.irs.gov/newsroom/irs-launches-new-effort-aimed-a...]
"The Internal Revenue Service today announced the agency has surpassed the $1 billion mark in collections from high-wealth taxpayers with past-due taxes. " [https://www.irs.gov/newsroom/irs-tops-1-billion-in-past-due-...]
--
Your ideological view of reality and what the IRS are objectively doing do not seem to remotely line up.
https://www.reuters.com/world/us/us-irs-prepares-lay-off-tho...
Except that Project 2025 wants to freeze the IRS budget and hiring:
> Doubling the IRS? The Inflation Reduction Act contains a radical $80 billion expansion of the IRS—enough to double the size of its workforce.34 Unless Congress reverses this policy, the IRS will become much more intrusive and impose still greater costs on the American people.
* https://static.project2025.org/2025_MandateForLeadership_CHA...
These are finanical systems, they need to add/subtract and multiply/divide lots of transactional data. AI not needed.
I (really) don't think AI is the solution in its current iteration, but it's unfortunate that it seems like they keep failing.
AI definitely won’t help with that.
https://www.theregister.com/2023/08/22/ibm_says_genai_can_co...
> IBM is giving its mainframe customers a tool infused with generative AI to translate COBOL code to Java
Hard to wholesale believe claims when they are literally the ones selling that capability.
It probably works well for already clear and well structured cobol.
It is one of the things they are best at.
Tho it works best with small chunks of well isolated code… I doubt it’d be very successful with the IRS’s ball of COBOL, especially considering how little cobol code is available for training sets.
It’d still require a ton of hairy manual work.
Plus tax laws change every year, in many cases causing massive changes. Every conversion failed because no one is left fully understands the whole system.
Plus for a conversion, you need a long quiet time, with the tax laws you never get that.
Not to mention, even with Biden's additional funding, which I am sure it will be cut, it is far from the amount needed for a successful project.
So here we are. I think some "modernization" has occurred, but the core system is still running some 60 year old software.
Note, I knew a lot of people who worked there prior to 2000, so maybe what I know is a bit out of date. But I doubt it since the attempts I have read about over the past 20 years did not succeed.
And while you're having a bunch of lawyers go through the tax code in order to do this, have them recommend simplifications to make it easier, so that instead writing new complicated software, you write new simpler tax laws.
If you have the gall to question IMF behavior vs printed tax code and write it up internally, you will get a reply from an tax code analyst in a couple of months from WV explaining why you’re wrong, selections from code, selections from IRS legal interpretations, applicable court cases past, present and pending.
It’s all very polite, but you quickly realize they’re lawyers and you aren’t.
People can still submit paper returns (and this is common for any businesses beyond a certain size- I've seen 200+ pages for a small-ish company's return). Supporting documentation and schedules can be non-standard (like receipts, letters, exhibits, spreadsheets, contracts, etc.), and the forms themselves can be hand-written and the other docs can be on inconsistent paper sizes so bulk ocr of any kind is rough.
Really, you have several challenging problems- maybe 200M individual returns filed every year, who knows how many corporate returns, all of the daily/monthly/quarterly filings for payroll, then for tax season you have an army of temp workers who are responsible for doing manual data entry which is cumbersome.
Then you have the financial management/reporting/reconciliation side of it. The volume is staggering. Electronic transfers in both directions, handling paper checks both to and from taxpayers. More manual data entry for manual payments.
And the kicker of all this is that it actually works for some pretty reasonable value of that term. And everyone here knows what the code probably looks like and we definitely know what the code's documentation looks like.
How would you like to sit down with the nearly 7,000 pages of the US Tax Code and start taking a bite of that monster? It's not a reference manual- the amount of cross referencing with other laws and regulations required would be monumental.
So, you are correct, AI is definitely not the answer; But I suspect the AI is for rooting out tax cheats (or maybe certain tax cheats), not running the system itself.
The IRS gets 125 million paper forms every year:
https://www.cnn.com/2024/01/29/politics/tax-season-irs-digit...
They tried and failed to scan some of them with OCR:
"OCR – This project involved extracting machine-readable data from low-resolution and poor-quality images (both digital and paper) with the capability to integrate them with the IRS’s information technology systems and intake different types of forms and data. The pilot project was discontinued in October 2022 after failing to identify acceptable solutions using the vendor and resources provided. The ED&CMO has no plans to resume this effort at this time."
https://www.tigta.gov/sites/default/files/reports/2023-09/20...
I imagine that a few GB200s would actually help with that effort
It doesn’t scale to 1040s and schedules because today’s biologics don’t know how to write anymore.
What made you think that?
So AI can be used to speed up transforming the existing infrastructure instead of AI doing the actual processing which would obviously be not the direct use case in anyones mind
It could also be implemented in less volumous tasks, like specialised scenarios like complex cases of fraud detection idk
Imagine how the appeal process will "work" once the IRS introduces LLMs for this...