Tech News of the Week 09-17-2024

0:08 Welcome to tech news of the week. This is our weekly tech news podcast where Chris and I get into four of the best tech news of the week. We're going to be doing a podcast. We're going to be doing a podcast. We're going to be doing a podcast. We're going to be doing a podcast where Chris and I get into four interesting articles that caught our attention. I'm going to go first, Chris, if that's okay with you. My mic's muted. Good. It should be. The EU takes a big bite out of Apple. The European court of justice has handed down a landmark ruling that forces Apple to pay 13 billion euros in back taxes based on what they judge. Based on what they judge to be an illegal tax structure put forth by Ireland. The case dates back to 2016 and regards a period of almost 11 years when Apple's effective tax burden in Ireland was a mere 1%.

0:56 Since that time, the loophole providing such relief has been closed and a universal 15% corporate tax has been adopted by most of the EU's member states. Not surprisingly, Tim Cook claims no wrongdoing, as does the Irish government. The case was originally ruled in Apple's favor back in 2020 by a lower court, but the European court of justice was primed to have the final say. And after a brisk four years, I think that's fast in the legal world. They decided that Ireland had granted Apple unlawful aid. You might be wondering about the money. Is the EU going to send Apple a PayPal invoice or a written proclamation by a carrier pigeon?

1:41 Will swallows, laden or unladen be involved somehow? Fortunately, no. The back taxes have been sitting in an escrow account since 2018. And with this judgment, they can finally be released to the Irish state. Just like with GDPR, the EU once again shows us the way forward on corporate taxation. If only we could get Amazon or Walmart to pay 15% of their net income with a lame duck president, like maybe he could do it. Well, or more likely Biden is too busy racing his Corvette down the Delmarva peninsula room, room bitches. Seems like a Camaro guy.

2:21 Maybe. Quantum update error corrected. Cubit count alert. As we've talked about a number of times on the show, creating cubits in quantum computers is getting pretty routine. I mean, relatively speaking, creating systems that can withstand errors, however, continues to be devilishly hard. Just so we're all on the same page, we are way over 1000 cubits in a number of running systems. So where are we with error corrected cubits?

3:02 You might ask. Well, Microsoft of all people announced an answer with what they're calling the largest current number of error corrected cubits. And that number is 12. The approach is interesting. Microsoft has partnered with a quantum computing organization called Adam Computing. The approach they're taking is to spread the value of each cubit across several cubits, thus making any errors or issues that come up, quote, less catastrophic. Hilarious language.

3:42 Love it. It looks like they're going with something around a four to one ratio, creating 12 logical error corrected cubits backed by 56 physical ones. And the approach does seem to be working at least for certain algorithms. The test improved the error rate from 2.4% down to 0.11%, which is substantial. Yeah. Now it's important to note that error corrected systems are helpful for a number of reasons, one of which is sometimes in quantum, there can be errors that can't be detected, which is different than errors that can be detected.

4:27 And I will leave the difference and challenge for both of them as an exercise to the reader. Long story short, though, spreading out the work and creating logical cubits like Microsoft and Adam are doing in this means that even these failures, the ones that are not detected can at least be mitigated. Neat. Open AI announces strawberry models. Quick open up chat GPT or copilot and ask it how many R's are in the word strawberry. Go ahead. I'll wait. Listen, buddy, I've got two liters of Joel Cola, a Sudoku book and adult diapers.

5:11 I can wait it out. You done? I could let us proceed before my heart leaps out of my body and strangles my teeth chances are your friend, I couldn't get through it. Chances are your friendly LLM told you that there are two R's in strawberry, which unless you are terrible at spelling, you know is wrong. So what? LLMs get stuff wrong all the time. Even better. If you tell it the correct answer, it will cheerfully suggest that you are the one counting stuff wrong. What is happening? It's like, I'm afraid you're mistaken.

5:53 There are only two R's. What is happening is that LLMs break things into tokens to process information and the word strawberry is broken into two separate tokens. The best guess is that chat GPT season R in each token and counts two R's. This thorny problem is so well known that open AI codenamed their new AI model line as strawberry, also known as O1 for reasons. The new model is allegedly capable of reasoning through an answer, much like a person does, instead of just trying to vomit the whole thing out at once.

6:35 O1 is the new model developed in parallel with the forthcoming GPT-5, and it makes use of reinforcement learning, aka telling the model when it gets things wrong. The reinforcement learning and multi-step reasoning should allow O1 to arrive at the correct answer of three for R's in strawberry, and also help it solve math word problems that have so far stumped previous generations. I got to try the O1 preview today and it apologized to me. Quote, "You are absolutely correct and I apologize for the oversight earlier. The word strawberry contains three R's." End quote.

7:16 Absolutely amazing stuff. AI dude bro lies about model capabilities. Gets caught. Hilarity ensues. This past two weeks has been pretty wild for other side AI. The company became AI world famous for its product, which is called Hyper Write, which is apparently a writing assistant. Is it hyper wrong? But um, but of course, success in the Hyper Write realm wasn't enough for other side AI. And thus they started hyping up their own AI. Going under the brand name reflection.

8:03 Allegedly based on llama 3.1. This past week, CEO Matt Schumer breathlessly announced reflection 70 B, which he claimed insane performance on. He showed tables and everything. He even published the model and uploaded it so other people could download it and test it. This is the first time that we've ever seen a model like this. This turned out to be a mistake as nobody could come close to the claimed performance numbers. In order to counter this, Matt went ahead and claimed that the upload was corrupted. Sure, Matt. Other side AI opened access to a private API so that people could test reflection 70 B at home base.

8:52 Seems like a not bad idea. Except that what the testers found was while there was better performance, there was plausible evidence that this private API was simply scrubbing answers pulled directly from an anthropics Claude model. Oh, so that's not a good look. After this Matt went dark, basically hanging all his supporters out to dry. Eventually he went on Twitter apologizing, sort of saying that he quote, got ahead of himself. This, as I'm sure you know, is also not a good look.

9:35 It turns out fake announcements of wild success using repeatable tests of known benchmarks against a product that other people can download is a bad idea. But Chris, he was in founder mode. Move fast and break stuff. We're done. Go away now. Bye.