"Perfect" is something that takes practice and refinement to happen reliably. It took quite a few launches before the Falcon 9 booster could reliable land too. There is a lot more scrutiny directed at Starship's refinement process than there was for Falcon 9. Falcon 9 was assumed to be expendable and successful at that. SpaceX started doing practice landings as a bonus activity. Sometimes it worked. Sometimes it didn't. But with progressive refinement it reached the point where nearly every launch ended with a successful landing and the whole equation changed. Starship's trials are viewed more sceptically because there is the notion that it HAS to recover everything in order to be successful.

Requiring a range doesn't mean much if the range is permitted to be inaccurate. Many published ranges ARE inaccurate. Candidates who been through the process for a while or multiple times know this.

I would generalise this and say that AI code is useful when it is easier to validate than it is to create. AI, as it is defined today, is an approximation engine that only asymptotically approaches reliable correctness. Thus, you must always assume that the output is wrong. Either being wrong doesn't matter of you must take the time to validate the result. My area is designing the logic that goes into chips. Coding effort is dwarfed by verification and debug. Correctness is critical. Once a chip is manufactured, there are no updates. And, yet, companies are now pushing us to use AI, too. A couple of weeks ago, I asked Gemini, Claude, Chatgpt, and xAI to generate a round robin arbiter and associated testbench. None of these models produced design code that would pass their own testbench. Gemini's code wasn't even a wrong round robin arbiter. It was some bizarre hybrid. And this is for a small, well known, foundational block. I can see using AI as a debugging too. If we already have a testbench that identifies the failure, It is easy to check to see if the proposed fix actually works. If it doesn't don't use it. But asking AI produce new, complex, and unfamiliar code seems like a dead loss.

Not even that. AGI doesn't have to be as good as a human at anything. AGI is not a level of intelligence. It is a different kind of intelligence than that the task specific pattern matchers we have today. It is kind of AI that we (collectively) don't know how to build at any level. An AGI with an IQ of 50 would still be a huge development.

Citation? It seems consistent with personal experience. I have only ever been poled while I was collecting unemployment. I have exhausted my benefits multiple times for extended periods, including now. Yet, I can't seem to find any official or even reputable source that says this.

Let me introduce you to the snowball effect. It is not necessary to be extraordinarily unlucky enough. It is only necessary to unlucky enough early in one's career that people assumed i, ii, or iv when seeing that you have a gap. this leads to relatively poor opportunities and in the course of time, more gaps and more poor opportunities. If the opportunities are poor enough (but you took them because you have to eat), they may fit your personality and strengths poorly leading to reduced performance too. Most engineers are poor at making connections anyway so that is nothing special but it does make it more difficult to climb out of a career hole. The lucky ones never have this skill tested and assume that anyone less fortunate is a poor hire.

Yes, it matters. Even if Apple's revenue depended entirely on hardware sales, relying on fewer sales at higher margins is a concern for future stability. Fewer sales suggest that Apple's products are becoming less attractive to consumers. Competitive pressures could force lower margins. But Apple also takes a big cut of App Store purchases. Fewer customers means fewer sales in the app store not just in the current quarter but going forward.

And no consumer mother boards until 2030? The spec was finalised in 2021! Gen *7* finalised almost a year ago. If I were still with Teledyne, I would be working on a Gen 8 analyser that would almost likely ship in early 2027, maybe even this year. It looks consumer motherboards are going to be four PCIe versions behind when they finally ship with Gen6.

The current need for humans in the loop is much less about performance and much more about accuracy. Amodei must believe (in public anyway) that the hallucination problem is nearly solved.

The non-graduates may be accepting low paying, low stability roles, while the graduates hold out for quality that has become scarce.

Sure, but definitely slower than a bicycle, which can easily manage 10mph in traffic. Being behind a bus is annoying as they are awkward to pass but you must pass them or be stuck at a reduced pace breathing diesel fumes.

Not directly in competition with AI. However, DRAM manufacturers are motivated to move production where it generates the most profit. If DDR5 is where the money is, expect production to increase at the expense of older lines.

So, are AI hallucinations as convincing as the lies told by human sales agents? This might be one of those areas where generative AI can actually be productive because the result doesn't have to be accurate.

A recruiter's only function is to help fill openings that the client has difficulty filling on their own. Of course, he thinks remote work is dead, those reqs are easy to fill and, thus, few of them pass his desk. The money is in hybrid and fully in office reqs. Those are hard to fill because the candidates don't want them. More work for the recruiters.