GPT-4.5 Drops & The Biggest Crypto Heist Ever | Tech News of the Week

[00:00:00.00]
Announcer: Welcome to Tech News of the Week with your host, a stool pigeon that prefers wingbacks.

[00:00:06.04]
Ned: Welcome to Tech News of the Week. This is our weekly Tech News podcast, where Chris and I dig into four interesting things that came across our desk in the past week. We have desks, right? Sort of. Yeah, we'll workshop it. I'll go first because you covered the big thing, Chris. Hashicorp is now part of IBM. Finally. Back in April of 2024, IBM announced their intention to purchase HashiCorp for $6. 4 billion, and now, 10 months later, the deal has finally closed. Originally, the deal was meant to close before the end of 2024, but both the US FTC and UK's Competition and Markets Authority dragged their feet in giving the acquisition the green light. They gave the go ahead at the beginning of last week, and within two days, the deal officially closed. Being an acquisition acquisition in limbo for 10 months is certainly no fun. Just ask VMware. Who hung in limbo for far longer. Both sides are breathing a sigh of relief that they can go back into full on planning mode and talk about the future in more concrete in concrete terms, something that is verbotim while the deal is still pending. What does that future look like?

[00:01:20.28]
Ned: Well, IBM's acquisition of Red Hat bolstered their aging product portfolio with cloud native application platforms and toolsets. And HashiCorp further fills out the picture of a more modern IBM. Terraform, the lingua franca of building cloud infrastructure, will be more closely tied to Ansible, which tends to handle the fiddlier bits of config management. Hashicorp vault is being integrated more tightly with OpenShift to provide for better secrets management than the vanilla Kubernetes options. I'm sure the other products, whatever they are, will find synergies in IBM's massive portfolio. If you'd like to hear more about the acquisition and the future of HashiCorp, check out a live stream I did with CTO and co founder, Arman Dadegar, earlier last week. Link is in the show notes, as they say. As a HashiCorp ambassador and fan of their toolset, I was not necessarily excited about the acquisition, but I've moved on to the acceptance phase and I'm pumped to see how much money IBM is dumping into their R&D efforts. Color me cautiously optimistic?

[00:02:34.10]
Chris: You wanted AI and Microsoft Notepad, right? Yeah, of course you did.

[00:02:39.06]
Ned: Sorry, I threw up in my mouth a little.

[00:02:42.25]
Chris: And lucky, lucky, Lucky you. Now you can have it. Notepad, the world's famous simplest possible text editing program that Microsoft ever built is now for some reason getting complicated. At one point, they added spell check. I was not okay with this, but fine. Fine? One program description that I particularly liked talked about Notepad as, It is a very limited editor with few features, and that's the way we like it. If we wanted more features, there's always WordPad. Oh, Wait, what? They're removing WordPad, a free program that's been part of Windows forever, and recommending people use Word, which you have to pay for. Well, that's shitty. But at least Notepad will stay pure. That's right. I forgot what we were talking about. Talking about how now they're adding AI features to Notepad that literally nobody could possibly have ever asked for. And of course, expecting you to have a subscription to M365 to use It. And they're pulling the same shit with Microsoft Paint. This is officially getting out of hand. This is not the point of these programs. These kinds of things have never been wanted or needed. We don't want anything remotely complex. The fact that paint and notepad are very limited with few features was the selling point.

[00:04:26.28]
Chris: Now, to be fair, they are both still available in their basic bitch models with the AI features grayed out for now. I might be old fashioned, but I think the very fact that we're using the word features in the same sentence as the word notepad is in itself a problem.

[00:04:47.24]
Ned: Agreed. So they added tabs to notepad, and I'm not mad about that.

[00:04:53.16]
Chris: So one thing. It's a slippery slope, Ned.

[00:04:57.26]
Ned: Gpt 4. 5 is bigger and better-er-er. We swear-er. When Deep Seq dropped a couple of weeks ago, side note, that was only three weeks ago. What is wrong with time? The three key takeaways that most people took are, one, there are far more efficient ways to develop and train a model. Two, general purpose doesn't scale and specialization is good, actually. Three, you don't need the latest and greatest hardware to make a capable LLM. So naturally, OpenAI has announced the availability of GPT 4.

[00:05:35.03]
Chris: 5, which is-They released this because they had to release something. Let's all be honest about it.

[00:05:41.16]
Ned: Christ. That's all I could think. Gpt 4. 5 is bigger than GPT 4. 0. It's completely general purpose, and it requires the latest hardware to run on, which OpenAI is now completely out of. Please send Sam Altman money or NVIDIA chips. Openai claims that the new model is friendlier and more truthful. Based on their own internally-developed simple QA benchmark, 4. 5 lies to you only 37. 1% of the time versus 59. 8 for 4. 0. Great. I've always wanted a friend who lied to me a third of the time, but was super chatty and disarming. Are we just building a con man? Yes. Yes, we are. Critics point out that GPT 4. 5 doesn't really improve over 4. 0 on the rest of the benchmarks, and the whole thing just reeks of desperation. After all, Sam Altman just moved OpenAI to be for profit and is seeking billions more in funding to prop up this colossal cash-burning machine. They could try to make a more efficient model that ran on the hardware they had, but while Altman's spell still holds sway over outfits It's like SoftBank, that seems unlikely to happen.

[00:07:07.19]
Chris: Cryptomarketbybit suffers the largest crypto heist of all time. Okay, So bear with me on this one as this is a rapidly evolving story, and I'm not convinced that I actually know what I'm talking about. I asked Ned for help. He, of course, gave no help.

[00:07:29.18]
Ned: I I gave you moral support. You're welcome.

[00:07:33.11]
Chris: Bybit is a company based in Dubai that trades crypto. It's a user platform like Coinbase, basically, operating an exchange where people can overspend on worthless Ponzi schemes one after another, depending on their tastes or the memes of the day. This week, it was announced that attackers from North Korea's Lazarus Group managed to break in, trick the UI, simulate real legitimate transactions, and walk out with $1. 5 billion worth of Ethereum tokens. This is the largest crypto theft to date, and depending on your accountant, is on par with the largest thefts in general of all time ever. This appears to have been possible because of the way that Bybit handled Ethereum. They allowed multi-signatures to manipulate an Ethereum wallet that otherwise should not have been accessible. And of course, as usual, social engineering wins again, as does convenience versus security. This model, it appears to have been built to make trading crypto, quote, faster. The fact that the system's UI could show faked transactions is concerning in and of itself. After the news of this came out, users of Bybit are removing money at a not unexpected pace. Bybit pledge to make all customers whole, and they have the revenue to actually to do it, which is good.

[00:09:08.19]
Chris: But yeah, long story short, this was called a cold wallet, which means it's not supposed to be accessible, and it's now an empty wallet, which just makes everyone sad.

[00:09:19.23]
Ned: Or it's just truly inaccessible because there's nothing there. Think about it.

[00:09:24.15]
Chris: Think about it.

[00:09:26.04]
Ned: All right, that's it. We're done now. Go away. Bye.