French Wine-making has no place in American Software development
Then, the French Wine industry froze it's self in amber and lost out to California. Now, California wants to do the same to it's own tech industry.
As with every single person who enters adulthood- you must pick a hobby that will consume most of your hours outside of work. Those hobbies are the ultimate proof of a well adjusted life. Mine just so happened to be the worst possible combination of hobbies- coffee and American legislative policy. My hobbies prove that I happen to be the most well adjusted in all the lands.
On the coffee front, friends of mine will tell you that I have fallen hard for specialty coffee. If allowed to speak independently, they will tell you that Angelo Saraceno makes one of the best pour overs they have ever tasted. They also will tell you how annoying I am whenever I find a new fixation and evangelize it till the ends of the earth. Outside of learning how to make a good cup, I have began researching producers, learning about harvesting and processing practices, and namely learning about the origins of the industry.
The one interesting thing that stands out to me is that the industry is progressing in the opposite direction of the software industry.
Software started off as a specialty industry where solutions were highly tailored to the needs of companies and are rapidly approaching commodification. (An essay for later.) The coffee industry is an industry that is trying to move away from wholesale bean trade to higher margin specialty coffee products. Recent innovations in fermentation, processing, and different plant cultivation is making it possible for the average consumer to notice and appreciate the taste of different origins and methods of coffee thus allowing producers to charge a deserved premium for their product.
In any industry that is going through a maturation process outside of hyper growth, usually the incumbents of an industry then attempt to play the referee. This happens quite often in the food and beverage industry. Usually a new entrant creates a new or rival category- the existing incumbents try their best to co-opt that new category or regulate it out of existence. (See: Oat Based Frozen Dessert)
This trend is now starting to play out over in tech.
As one phase of hyper growth ended, the incumbents of a burgeoning industry in A.I. and hardware hope to front run the regulations. A rather reprehensible bill primarily championed by California State Senator Scott Weiner over in California known as SB 1047 just got passed in the state legislature. It is an A.I. safety bill that aims to use tort liability as a way to enforce A.I. safety. However, proponents of the bill likely real aim of the legislation is to "pause" development of A.I. Rather insidiously, the same panel of newspapers who have fawned over FTX are on the same side of this legislation unfairly portraying the proponents of this bill as "Big Tech"- when it aims to expose developers to civil tort liability that is more than unreasonable.
This is not hyperbole. It aims to add onerous regulations and liability for a technology that frankly, doesn't exist. A.I. isn't sentient, it is an advanced next word predictor.
If someone uses a "Covered model" in a harmful way the model developer is exposed to the liability of the use of the model or operational leaks. If the weights of the model are leaked by a hostile actor, like Meta's Llama3, well, Meta is on the hook for an exorbitant fine and civil suits. You can read the bill yourself, it's quite understandable to lay people without having to be familiar with tort law- and it's bad.
Although later revisions of the bill watered down the the initial legal liability of harm done by frontier models- it does by primarily by establishing a 9 person panel to be the final arbiter on whether a company has the right to develop anything. This is the latest continuation of California's continual desire to be a nanny state that seeks to regulate innovators while ignoring other legislative priorities. Like most laws that aim to use the legal system to protect a imaginary constituency, it hampers innovation while ignoring the global nature of software. As far as legislation goes, there will be a lot of negative externalities. The chief one being is the assumption that frontier A.I. development will stay in California.
Believe it or not, this is not the first time that something of this sort has happened- ironically it involves California. An established wine-making industry in France faced a new challenge from upstart farmers in California with a superior product. What did French producers do? What even does wine-making, coffee, have to do with AI? We'll tie those knots and more, but first some context and history.
The Upmarket Cheese Tactic
I have been avidly following the journey of coffee roasters and producers as they harvest and refine their produce to a consumable end product. Much like how inside baseball in the software industry usually revolves around: "should we have software estimates?", "dae agile?" - the coffee industry is undergoing a few debates between major and upstart producers and their farmers. Namely, the big debate is- how do they move their product upmarket?
The only problem is that due to a number of factors, farming specialty coffee is extremely time intensive, requires cultivation of certain species of plants, and requires specific expertise that not all farmers have access to. To shortcut those factors to make it easily marketable for consumers to know that they are getting a quality cup, usually a roaster will market their bag of coffee as single-origin to note that their coffee is of superior quality. In the past, this has been a reliable indicator for consumers.
However, what's been happening is that some producers have been cargo-culting the intent of single-origin coffee to signal that it is of quality and confusing it with the notion that "only coffee that tastes this way can come from this farm/place". The chief violator of this idea is Nestle and Starbucks.
Side note: I assert that those producers actually aren't even coffee enthusiasts in the first place where the management level is staffed by a cadre of McKinsey B-team hires that only recently are starting to get cleaned from the company. (The Onyx Coffee Labs, Black/White, and September Coffee's of the world aren't pushing this stuff.)
That is harmful marketing, especially when it's not true.
Now, a reliable signal in the past is turning into noise as larger producers are taking note of this and marketing coffee from a single origin with lower quality beans. (This phenomenon also happened to the much vaunted fair-trade label in the past, but that's a story for later.)
As coffee producers started making a differentiated product, they realized that one of the quickest ways to defend that product offering is by claiming that they are the sole providers of that experience. At best, you do that by offering an excellent service, at worst you do that by monopolizing. This is exactly what happened in France... on purpose and then by accident.
A Misplaced Notion about "Terrior"
Around 2018, the coffee industry saw the status-quo of hundred dollar bottles in the wine industry, and said, "I want that"- and then proceeded to speed-run all of the mistakes of the wine industry while ignoring the history and the context of those choices.
All in the pursuit of: "Terrior".
I knew about this concept because while in Spain, I learned about this association that was referenced by a wine maker who was lamenting that his wine doesn't sell because it wasn't French. As a red blooded capitalist, that made me sad to hear, his stuff was genuinely good. Luckily, while in my coffee kick- I found out about Lucia Solis' work in coffee processing who's been working to debunk that notion while in the coffee industry.
What is "terrior"?
Roughly translated... it means soil. The French Wine industry are master marketers using their unique blend of heritage, experience, and... lies to sell differentiated experiences that buyers around the world are willing to pay thousands of dollars per bottle for. The chief notion behind that defensible value is the notion of "terrior". If you went to a wine enthusiast and asked them where would the best tasting wine would be found- they would name a few vineyards across Europe and America. However, a casual drinker would possibly reference, France. It is that association, assisted by government intervention that reinforces that notion.
The coffee industry, for a moment, wanted the same level of marketing cache that it was the unique properties of the land of a region that it was grown in that gave a coffee a certain taste. This ignored the long march of innovation that comes for everyone. ...While also ignoring the reason why wine went from everyday consumption good to luxury good.
See, the great wine estates in France were and are helmed by businessmen who became winemakers rather than becoming wealthy by wine-making. Example, like in 1570 where Pierre de Lestonnac in Margot got rich first and then bought the lands that would eventually become Chateau Margeaux. Other houses would form similarly to this system. After a period of what can be summarized as: self promotion and patronage, they would establish themselves as the premier wine-houses. However, it was the prestige of the families, not the wine, that would lend credence to the wine's reputation.
This would be solidified in 1855 under Napoleon III, during the 1855 Exposition Universelle where the great houses featured their wares and were classified by the fair, not for quality or feedback, but only to differentiate for fair goers, which wines were quality. Imagine taking a test one time and being evaluated by that for nearly the remainder of your existence. That was how French houses were graded- once. This would create a ranking system that to this day is functionally immutable despite the the varying fortunes of the houses producing inferior wine- such as the time when the economy of Bordeaux was in crisis in 1973 with Chateau Margeaux entering administration.
In parallel, in 1976, the Californian agricultural economy was booming and to boot, they were producing a number of new farms and wines that were starting to get global attention. (Including with a house that was an offshoot of a fifth growth French wine maker, I forget which.) Steven Spurrier, a British wine merchant wanted to put this to test and see if California wines were any good. He set up a blind taste test with American and French judges in the panel and... a Napa valley wine won at what became called The Judgement of Paris. Funnily enough, many California wineries that were formed which won the taste test were founded by entrepreneurs who got into wine as a hobby after amassing their fortunes in the gold rush. Fun fact: Alfred Tubbs who founded Chateau Montelena one of the houses of the willing wine made his fortune by selling candles and rope.
Lucia Solis attributes the wins in California to innovations in farming practices and cultivation. Napa Valley's reputation was cultivated by merit, not by decree.
Back to coffee... Solis mentioned here moved from the wine industry over to the coffee production industry due to her frustrations with the hierarchical nature of the wine industry. Although she isn't the pioneer (or she's too humble) of newer methods of coffee processing. Around 2019, a development started happening with regard to the taste of single origin coffees... the taste of the crop could be manipulated by natural methods. Thanks to advances in washing, roasting, and fermentation, it is now possible to get Colombian arabica beans to have the florals and juicy tasting notes of an Ethiopian origin coffee. If the coffee industry kept on running head first into marketing single origins as much as they did they would have ran into issues when these new innovations came along.
Today, in specialty coffee, now process, the farmer themselves, and taste notes have been key signals of quality and taste over the origin and certifications that are now falling out of favor in the coffee industry.
What does this have to do with A.I. and software?
In my parallel- I am finding that the software industry is all too eager to invite the referee in hopes to solidify the status quo of big tech domination. As someone who works for a startup with meaningful ambition, although I am not personally affected by the latest salvo of regulation, I am concerned with the continual eagerness by think tanks and legislatures alike who are empowering the powerful who can afford lawyers and the person power to deal with the laws that are coming down the pipe.
To be clear... personal opinion, but take SOC2 for instance. This is fine as a industry self-certification, but I would voice vociferous opposition if it was mandated by the State of California for every single company who operates there. That's effectively a three month regulatory penalty on a startup’s cycles. A reasonable certification but a unreasonable ask to move a company's focus from it’s responsibility to it's product and operations.
The lesson I take from the escapades of the coffee industry trying to develop a strong link between soil and the perceived quality of a product is that California legislators are trying to enforce that bond between the research labs and companies primarily located there and the state, making the fatal assumption that the tech industry would continue to exist there. (This is ignoring their tax law changes too.) Specifically for S.B. 1047- it makes assumptions about the state of compute and models that ignore the rate of progress that happens in *any* industry. Take for instance the provision on covered models.
(A) Before January 1, 2027, “covered model” means either of the following:
(i) An artificial intelligence model trained using a quantity of computing power greater than 10^26 integer or floating-point operations, the cost of which exceeds one hundred million dollars ($100,000,000) when calculated using the average market prices of cloud compute at the start of training as reasonably assessed by the developer.
(ii) An artificial intelligence model created by fine-tuning a covered model using a quantity of computing power equal to or greater than three times 10^25 integer or floating-point operations, the cost of which, as reasonably assessed by the developer, exceeds ten million dollars ($10,000,000) if calculated using the average market price of cloud compute at the start of fine-tuning.
(B) (i) Except as provided in clause (ii), on and after January 1, 2027, “covered model” means any of the following:
(I) An artificial intelligence model trained using a quantity of computing power determined by the Government Operations Agency pursuant to Section 11547.6 of the Government Code, the cost of which exceeds one hundred million dollars ($100,000,000) when calculated using the average market price of cloud compute at the start of training as reasonably assessed by the developer.
(II) An artificial intelligence model created by fine-tuning a covered model using a quantity of computing power that exceeds a threshold determined by the Government Operations Agency, the cost of which, as reasonably assessed by the developer, exceeds ten million dollars ($10,000,000) if calculated using the average market price of cloud compute at the start of fine-tuning.
Two years ago, we weren't even talking about "GPT" the way we are now- enshrining this limit today without regard to progress means that it locks in the players we have today who can get access to this compute. This limit effectively declares first and second growths in regard to AI companies who can afford to deal with the additional procedures that a company would need to submit for audit. One of the bill's supporters Calvin Nelson mentions that this is a reasonable limit and uses existing tort provisions, but this will increase further liability for development of future models. If I was on the legal team at OpenAI, I would reconsider my presence in California because I accept this graph as law, not conjecture.
Furthermore, the tech industry wasn't always a Californian mainstay.
It used to exist in Menlo Park, New Jersey that is...
...where Thomas Edison pursued a commercially viable incandescent light bulb.
(Ironically, the Menlo Park in New Jersey was named after a failed Californian real estate development.)
That development gave us, for better or worse, 'Ma Bell, but gave us Bell Labs, the most important applied research institution of all time. It was the non-enforcement of non-competes in California plus their assortment of national labs that eventually gave rise to it’s industry. Their (healthy) disrespect of the law within the country gave it a gift.
At the moment, through a variety of pain pricks applied by the state, it seems that California politicians are hell bent on scaring away the golden goose that has been the tech industry. Local Californian legislatures are intent on making life there unsafe by not prosecuting petty crime. State legislatures are intent on punishing companies by regulating research activities. I agree with the spirit, it's important to hold power to account, but the legislation and it's amendments are co-written by companies like Anthropic that seek to freeze the status quo. (Primarily by limiting access to training frontier models stateside in California where most GPU hardware are located... for now.)
But it’s a own goal in the name of safety.
Safety, but for whom?
But the limits are naive thinking and will have the opposite effect especially by the bill's proponents.
At the start of the essay I said:
I assert that those producers actually aren't even coffee enthusiasts in the first place such as Starbucks which at the management level is staffed by a cadre of McKinsey B-team hires that only recently are starting to get cleaned from the company. (The Onyx Coffee Labs, Black/White, and September Coffee's of the world aren't pushing this stuff.)
When looking at the bench of proponents pushing the bill- I found something interesting.
The AIPI is headed by Daniel Colson, who is closely associated with Samo Burja. That group of safetyists are headed by rationalist over-thinkers who imagine doom and seek to put America in a form of permanent Luddite state even if that meant that hostile state actors or non-state actors feast on the end trails of our country's demise. (Elon Musk is part of that group.) I am wholly confident that these people have never trained a model in their life. The United States and it's penchant of rule by non-practitioners never ceases.
Do we want a next token predictor making a decision about the nations nuclear arsenal?
No.
Can that be done today with IBM's expensive gov't contracting work today?
Yes.
To me, it then means that training guardrails are the wrong way to apply the guardrails of safety that generative technologies can perform. (Nor should that be the reason to expose any company to civil tort liability.) There exists enormous risks that we are already facing, take deepfakes. I believe we don't have sufficient laws to prosecute cyber-crime: today. I would have wholeheartedly supported criminal provisions for those who engage in artificially created revenge porn. I would have been much more enthusiastic to read about anti-fraud measures to put in place to prevent impersonation like someone calling a bank on my behalf with my voice... not whatever SB1047 is which would effectively put Meta on the hook if any of the above actions occurred. As it stands today, this bill would even fail to protect the most vulnerable among us who are most at risk of this technology. Instead S.B. 1047 seeks to place blame on the researcher.
Working the Referee
Look- I am not a CATO free-markets, free-borders type of guy. I like the FDA, they get a lot of undue hate- I like the fact that I can take Advil and know that the regulations applied today means that I won't become comatose by taking a pill. Regulations are needed to prevent charlatans from selling snake oil, because if given the opportunity to do so- someone out there will.
However, the source of support of SB1047, the LA Writer's Guild, the AI Safetyists who are put out of a job because GPT-4o can write a better Lesswrong post, or anyone who fears for their job because they assume that their work will be automated- are mistaken. This is supposed to be a "feel good" bill that is red meat to labor. But, whenever when any innovation happens, 9 out of 10 times, any attempts to regulate automation on how work is done leads to the demise of an industry in that country. Namely, due to the labor costs of that work increasing to the point where it's wholly cost ineffective to continue that industry in that location.
It happened with steel workers in America losing out to Japan embracing automation.
It happened to heavy industry in America moving to South Korea embracing automation.
It happened to the car industry in the United States leaving to Mexico and then now China.
It's happening to the animation industry where Indonesian animators are producing good quality work with technical assistance.
...and it will happen to software if we aren't careful.
I hope I am not correct in expecting the United States will enter a state of stasis much like the one our European friends are facing due to the onslaught of regulation that our industry is facing. What would be ironic is that it could die at the hands of non-profits and think tanks funding time billionaires to waste the time of innovators having to debate bunk bills like these. More ironic is if France succeeds in creating the conditions where a next generation A.I. company can prevail.
It would be quite a shame if California became only known for wine and and the coming graveyard of companies thanks to poor policy like this.