I don't know if it's the same in the US, but every job in Australia generally comes with a 3-6 month probationary period where employment can be terminated for basically any (non-protected) reason with little notice. I've observed that the option to do so is seldom used - probably because there is not usually much incentives for managers to decrease their number of reports unless there's serious problems.

Most places still interview. Because hiring someone for 3-6 months is still an expense. For large companies, onboarding can take many weeks before you're even thinking about being productive. Interviews don't need to be 100% accurate. They need to be time-efficient and an ok filter to prevent hiring the worst people.

I don't see how a bad job market is going to make skilled people be willing to intern in order to get employment. Employment is a market, if demand goes down, then the reaction will be that price goes down.

The reason students are willing to intern is because the supply is so high and demand so low that you can effectively hire them for nothing (or next to nothing). The interns know that on completion the internship will upgrade them to a new category with new supply/demand. It's the same reason they are willing to pay large amounts for a university degree.

So interviewing will stay the same. If demand collapses, we will see wages drop, and I imagine the supply of software people will react through early retirement, career switching and reduction of people choosing it as a career before we see an upheaval of hiring methods.

That probation period never ends in the us. You can be fired at any time for nearly any reason.

Or no reason.

Well... in normal times they would be entering at the bottom of the index due to the company beginning to grow, the purchase of which is being funded by a firm exiting the index due to shrinking, so assuming you have bought and held units in the fund, most of the time an index fund is buying low and selling low.

And then when you sell your units, hopefully in aggregate the index is worth more than it was when you entered...

First sigmoid was transformers allow us to rapidly scale to our already abundant data until we tapped it out, the second is/was reasoning, allowing us to scale to our available compute (and compute manufacturing capacity). Correct me if I'm wrong but we don't have candidate for the third sigmoid, and scaling inference is hitting real-world supply chain constraints - electricity and chips.

Short of a third sigmoid appearing in the ML CompSci space, perhaps in the form of ongoing, repeated step-optimisations which will also have diminishing returns, intelligence growth is now limited a few scaling problems that have already been worked on for a very long time.

Transistors, which have been doubling for almost a century now, but Moores Law has already plateaued and reached limits on energy efficiency, and simply building new fabs is not something that we can do exponentially. And the other growth limiter is electricity - there is no exponential supply of fossil fuels or power plants. Although manufacturing has scaled, PV tech improvements are also plateauing - and while storage is getting cheaper, it's still not economical vs fossil fuels (meaning: when we have to switch to it, the growth slows down further) and we are unlikely to see battery efficiency sigmoid enough to maintain the AI sigmoid.

I don't mean to be bearish here. There's so much money sloshing around that we can afford to put the smartest people, using unlimited tokens, on the task of finding small, incremental gains on the CompSci side of things that will have large monetary payoffs - hopefully allowing further scaling and increased emergent abilities of LLMs. Maybe we can squeeze the algos for quite a while. But I don't see that maintaining the same level of exponential as unlocking unlimited data or maxxing out the world's energy/fab capacity for long.

And I don't see why this is a massive issue except for the people who want to have some god-like super AI? Frontier LLMs are genuinely magic. Not "won't delete your production database" magic, but definitely a massive productivity gain for competent knowledge workers.

We are multiple orders of magnitude away from Landauer limits - so next big thing in matmul could be photonic multipliers - there’s a bunch of them coming up in the next 3? years. So that’s a 2-4 order of magnitude improvement. Sigmoid?

Reinforcement learning has become a huge portion of compute used during training runs [1] and synthetic data is letting us get lots more mileage out of the existing data. Additionally, there is lots of new, high quality data being created and collected each day. I think the "running out of data" thing was pretty poorly reported by mainstream media.

[1]. https://www.dwarkesh.com/p/dario-amodei-2

It is for now.

But I'm sure the scanning operations will start scouring the earth even harder for any books unaffected by slop containing niche knowledge and text in order for their models to have an edge over the ones trained only on pirate collections and the Internet.

I wonder if secondhand bookshops and deceased estates are seeing bulk buyers of their stock suddenly appearing. Maybe broke governments/municipalities will start selling them entire libraries and archives to ingest.

> I find it hard to believe that the knowledge to manage a bunch of dedicated servers is that arcane that people wouldn't choose it for this kind of gigantic saving.

Managing servers is fine. Managing servers well is hard for the average person. Many hand-rolled hosting setups I've encountered includes fun gems such as:

- undocumented config drift.

- one unit of availability (downtime required for offline upgrades, resizing or maintenance)

- very out of date OS/libraries (usually due to the first two issues)

- generally awful security configurations. The easiest configuration being open ports for SSH and/or database connections, which probably have passwords (if they didn't you'd immediately be pwned)

Cloud architecture might be annoying and complex for many use-cases, but if you've ever been the person who had to pick up someone else's "pet" and start making changes or just maintaining it you'll know why the it can be nice to have cloud arch put some of their constraints on how infra is provisioned and be willing to pay for it.

For the record, I have seen every one of those in cloud based hosting multiple times. None of those issues require special work any more than they do than in traditional hosting.

I cannot reconcile that growth for non-technical users is going to explode, when most utility from agents is via the ability to execute arbitrary code, generally in yolo mode, with the fact that almost all corporate IT departments do not give users the ability to install anything on their machine, let alone arbitrary code. Even developers at many companies are subject to this despite the productivity impacts.

The culture of corporate IT would need to change to allow it, and I just don't see it happening.

Someone correct me if I'm wrong, but an LLM does not interpret structured content like JSON. Everything is fed into the machine as tokens, even JSON. So your structure that says "human says foo" and "computer says bar" is not deterministically interpreted by the LLM as logical statements but as a sequence of tokens. And when the context contains a LOT of those sequences, especially further "back" in the window then that is where this "confusion" occurs.

I don't think the problem here is about a bug in Claude Code. It's an inherit property of LLMs that context further back in the window has less impact on future tokens.

Like all the other undesirable aspects of LLMs, maybe this gets "fixed" in CC by trying to get the LLM to RAG their own conversation history instead of relying on it recalling who said what from context. But you can never "fix" LLMs being a next token generator... because that is what they are.

I think that’s correct. There seems to be a lot of fundamental limitations that have been “fixed” through a boatload of reinforcement learning.

But that doesn’t make them go away, it just makes them less glaring.

That's exactly my understanding as well. This is, essentially, the LLM hallucinating user messages nested inside its outputs. FWIWI I've seen Gemini do this frequently (especially on long agent loops).

My favourite library from these folks is gum (https://github.com/charmbracelet/gum). The basic premise is simple - instead of using hardcoded variables or in addition/instead of using CLI flags, call gum and capture the STDOUT to get the selected input value(s). Great for turning a bash script into a TUI, uses these libraries under the hood.

I find the pattern of showing interactive TUI if required options/flags are omitted much nicer than showing an error/help output.

Future models know it now, assuming they suck in mastodon and/or hacker news.

Although I don't think they actually "know" it. This particular trick question will be in the bank just like the seahorse emoji or how many Rs in strawberry. Did they start reasoning and generalising better or did the publishing of the "trick" and the discourse around it paper over the gap?

I wonder if in the future we will trade these AI tells like 0days, keeping them secret so they don't get patched out at the next model update.

The answer can be “both”.

They won’t get this specific question wrong again; but also they generalise, once they have sufficient examples. Patching out a single failure doesn’t do it. Patch out ten equivalent ones, and the eleventh doesn’t happen.

Yeah, the interpolation works if there are enough close examples around it. Problem is that the dimensionality of the space you are trying to interpolate in is so incomprehensibly big that even training on all of the internet, you are always going to have stuff that just doesn't have samples close by.