It's good to see honnibal staying engaged!
4,247 karma · joined June 2, 2009
Are the features designed well (from a product perspective) but implemented poorly? Again, the process is dysfunctional to let them get this far before it notices.
Were there known issues with these that were planned to be fix, but the fixes didn't happen in time? That would make sense and it happens, but how is an LLM making that worse? LLMs in the worst case let you get the same work done more quickly, because you can always get the work done as quickly as you ever could by just not using an LLM. So the strongest thing you might say here is that you thought LLMs would let you get a feature into shape before it was too late, and for whatever reason that was overly optimistic.
https://www.postgresql.org/message-id/20140818195634.2543.56...
It seems quite possible that someone wants to make it seem like LLMs are causing problems by over-reacting to things that could just be fixed, or delay the release slightly (or maybe waiting until late before a release to kill features they don't like?) Whatever it is, this seems like a dysfunctional process.
I think we can let him take a little credit for the work he did without trying to exactly determine how much of that credit is his exactly.
The yield of this wheat is 1.5 (Miracle Wheat) vs 5.8 (Standard) tonnes/hectare. That's about 25%. Since wheat can easily be stored for many years, you would have to have droughts in 3/4 of years before you break even due to drought resistance. In the meantime farmers can barely break even financially growing this stuff.
*HOWEVER* Miracle Wheat is very very susceptible to lodging (falling over and dying in the rain). So in years where there isn't a drought you would sometimes lose almost 100% of the crop. So even if there were droughts 3/4 of the time, you would still never break even.
There are numbers for this specific wheat. Emilia-Romagna's 2016 agricultural economic assessment compares grano del miracolo with conventional wheat [1]:
Yield: 1.5 vs 5.8 tonnes/hectare.
Variable costs: €527 vs €570/hectare.
Selling price: €390 vs €203/tonne.
Gross margin: €58 vs €607/hectare.
Preserving genetic diversity is valuable, and specialty buyers may pay enough to make growing it worthwhile. But that specialty market is precisely what I mean by "novelty." The economics of growing it as ordinary commodity wheat are poor.
[1] https://agricoltura.regione.emilia-romagna.it/psr-2014-2020/...
Statistically there is still room for there to be some unexplained physical reaction to gluten that is not explained by Celiac. However, nonexistence of such a reaction is also compatible with the studies. What is not compatible with the studies is more than a single digit percentage of non Celiac having people, as a % of people who claim to be gluten sensitive, being able to even detect whether they are fed gluten vs a placebo, despite it being very trendy to claim that you are 'gluten sensitive'.
People who do not have Celiac and say they are gluten sensitive usually are not. They are either experiencing a nocebo effect or have gas from other components of wheat rather than gluten. Actual non Celiac reactions to gluten are minor if they exist at all. So when someone says they are gluten sensitive it is overwhelmingly likely they either have Celiac and even a very very tiny amount of gluten will cause them horrible consequences, or they are incorrect and they are not sensitive to gluten at all.
This grain, while interesting, is far far less productive than modern varieties and less efficient to grow. It's valuable for genetic diversity and rediscovering it is wonderful and valuable, but it is more or less worthless for farmers as anything other than a novelty.
The benchmarks often have ambiguous instructions, and there is no interaction with the agent. So a docker image is started, the agent is called in non-interactive mode with a prompt saying to read the instructions and finish the task. The instructions are often somewhat ambiguous and the agent has no way to ask for clarification. If it picks the wrong option out of many seemingly equally likely options that test will be mark as failed.
Here is your first example, "corrected": https://share.gemini.google/DCSLT4PPap5q
Having an AI do a quick check of research studies is about as effective as just trusting a Vogue magazine article. You are just fooling yourself into thinking there is something useful there, there is not.
You are assuming that Carbon is ready for release, deployed to their datacenters, that all scaling and other issues have been worked out, and they are confident in all of the above. Not to mention additional internal constraints like alignment testing or other safety procedures their own policies and guardrails require.
If they are smart, which they clearly have some smart people if they are making the best frontier model, they would account for the above needs in their timeline, and announce the model when they can prove they have it but also before they are ready to release it. This is so they can be at the top of the frontier model lists, boost their brand for both Gemini and Google, get the bump in stock price and investor confidence, steer adoption decisions for people making decisions for what platform they might adopt, etc.
Windows 11 is awful, lots of the fixes and modernization started in windows 10 has totally stalled, leaving users with 5 different places to change a setting instead of the one old place or one new place. Look how many different UI's there are to adjust sound card / sound settings or display settings. It's a complete mess. Their attempts to jump into the AI game have been a complete joke, despite already having integrated Cortana years ago. Putting Cortana on an llm is such an obvious thing to do, and they already did so much work to make that fully integrated into windows and force it on everyone even though it was trash, and now it's still trash and they don't seem to be able to adapt at all.
Now they can't respond to a complete insecurity in one of their major products in weeks? Why do they even validate certs? Why doesn't revocation work in their app? I don't think they can really be trusted at this point.
I said that their strategy is self defeating. You are arguing they will do it anyway.
You are arguing that people are delusional about their worth. My point agrees with that, most people are, but there are some people that are actually valuable and companies are very bad at telling the difference.
It is not true if you are determining the correct feature to add to microsoft's money making products. That requires deep knowledge of users and the industry and the existing product. It requires understanding the industry and users far in advance, the turnaround time for a feature in Windows or Azure is likely 6 months or longer, sometimes years. It requires acumen and an ability to think about customers and the market and sales. Microsoft doesn't have 5000 people who can do this. They probably don't have 1000. These people can go start their own business or consulting firm today. There are lots of people they worked with in the past who know how good they are so they can get another job.
The same is true for engineering. Out of 500 engineers, there are 10-20 that are doing the really important work that you can't just find someone else to do. They know everything about the system and everyone comes to them for advice or to get explanations about how things are put together and why. You can't just find some average engineer and put them in the same desk as these 10-20 and magically things are fine, the mid tier person doesn't have the years of experience nor the potential to fill this role even if they did have the experience.
Shooting someone is an intentional act. Setting up a booby trap to shoot intruders is automating something that you could have done legally (shoot an intruder) except there is NOT intent. Intent to shoot the intruder is what makes it legal, so long as you meet the other requirements for self defense or the castle doctrine. Shooting someone without intending to shoot someone is negligence or worse. Building a contraption to shoot someone is setting in motion a chain of events specifically to cause a person to be shot with no possibility of satisfying the requirements to justify it.
Here's another scenario: you are cleaning your gun, when unknown to you a home invader is preparing to kick your front door in. Your gun goes off accidentally, shooting the home invader before he gets to your door. Legally it is not self defense, you have negligently shot someone. However, the robber can't sue you as felons do not have a recourse for injuries proximately caused by their own felony (if they are convicted) in California. In other states, the robber may be able to successfully sue even if he is convicted. Laws are weird.