If you, tersely, ask google “length of a 48 bit number in decimal” it comes back with
"A 48-bit number requires exactly 14 digits to be represented in decimal form.
The maximum possible value for an unsigned 48-bit number is (2^{48} - 1), which equals 281,474,976,710,655."
Is you ask a bit more verbosely you can get
“This means the decimal representation will be 14 to 15 digits long”
Wonder what that does for people who expect it’s answers to be ‘real’?
Answers are never “real”, or logically analysed, or even “correct” from an LLM. That’s not how they work, now or ever. It’s just “pattern matching” and “token prediction”. Very good at that, but that’s all it is. Somehow 99% of the population think that AI as it currently stands is a “logical” model and it “understands” the question. Never is and never will be. It’s not even deterministic. Good luck outsourcing decisions to AI with no human oversight…
PS I think AI is a “blight” on the world (a net negative), but that’s another issue (and would take me pages to explain why I think that way) 
Cheers, Misha
So, what ‘marketing’ currently calls AI, isn’t even the beginnings of a true AI, got it.
It’s just an extension of Eliza, from the 1960s, wonder how much of the code is unchanged.
So, the results we get at the moment is more of a tribute to, at least for Delphi, the helpfulness and correctness of Peter Below (and others).
So, if their correct ’cut’ was added to all the electricity and data centre costs no one would use it at all.
AI via LLMs have made huge leaps forward, but their benefits and limitations just need to be understood properly (and they aren’t). Three major breakthroughs in the last decade: Use of the transformer to “categorise”; access to the world’s knowledge; and an ability to process on GPUs. But “intelligent” by our definition they aren’t. I’m happy now with the term “artificial intelligence”, but it isn’t the same as “human intelligence”, and it’s nowhere near “general intelligence” (and never will be using LLMs)
I use sparingly (I use it as a “better Google”), but anything I use from it either A) doesn’t matter if I get it wrong, or B) if it does matter, I check everything from original sources and use my judgement on all decisions. I never outsource decision making to AI (why would you when it is just a big pattern-matcher?)
In terms of development, I write as little code as possible, use a very small number of established frameworks, and think about things for a week before coding. A good day for me is when I remove code. I work for myself (and directly for two clients), and my output is measured by whether something works, not number of lines of code. I don’t use AI-assisted coding in any way because it just won’t help me. If I totalled up the amount of debugged code that goes into production I develop in a day, most would think I’m on holiday 
This system (https://app.brook.com.au/racingconnex) is less than 10,000 lines in total (you can’t see the pages, but they allow stewards and judges to view and officiate races, power kiosks at the track, and supply overlays for Sky racing). The system also has live AWS messaging for the TAB and State racing bodies, a generic API, and an SQL Server DB. If I can do that in under 10K lines of code, why is everyone writing so much code?
One of the directors in a company I’m chairman of always uses AI of a company we might invest in. Some of it is almost always wrong.
Because that seems to be the ‘modern’ way.
For one of the things we do we ask for ‘sample code’ for any hardware we need to support (because the ‘documentation’ never covers the ‘quirks’, and if you don’t code for the quirks, well it never works), and then rewrite it for our ‘system’, the sample code these days is generally 20-30MB zipped, we battle our way through it and end up implementing it in a couple of hundred lines on our system. Or to put it another way, the sample code compiled is often 4 or more times the size of our entire system, and yet once implemented in our system it is way less than 1% of it.
You need to be very careful to keep up-to-date with the latest models before having an opinion. Its changing faster than any tech I’ve seen in 25 years.
For example, what’s wrong with ChatGPT’s answer?
A 48-bit unsigned number ranges from:
0 to 2^48 - 1
2^48 - 1 = 281,474,976,710,655
So the maximum decimal length is:
15 decimal digits
For signed 48-bit:
-2^47 to 2^47 - 1
Max positive:
140,737,488,355,327
Also 15 digits, or 16 characters including the minus sign.
OR Googles,
The length of a 48-bit number in decimal is 15 digits for the maximum possible value, as the number of digits ranges between 1 and 15 digits.
Detailed breakdown:
Maximum value: \(2^{48} - 1 = 281,474,976,710,655\) (15 digits)
Minimum value: \(0\) (1 digit)
Total unique values: \(2^{48} = 281,474,976,710,656\) [1]
You can calculate the maximum length of any binary number in decimal by taking the base-10 logarithm of \(2^{\text{bits}}\) and rounding up:
\(\log_{10}(2^{48}) = 48 \times \log_{10}(2) \approx 14.449\)
Since the number of digits must be a whole number, rounding \(14.449\) up means the maximum value will span 15 decimal places.
AI has changed development forever, is smarter and faster than most developers especially if you consider the response time and context available instantly, and the latest models perform in the top 1% of all developers out there on general work. It is true you might not be seeing this against Delphi code, but this is more an availability of code issue, not a model issue.
If you’re not using the latest tools you may not see these types of results, but this isn’t the kind of technological change where you can use that traditional ‘pop in last year, take a look, not be satisfied, mark it as rubbish forever’.
I’m not about to say AI coding assistance isn’t great (if you write a lot of code), but if what you are mainly doing is coding, you aren’t likely a “software developer”, and definitely not a “software engineer”. AI isn’t going to replace a true “software developer”, and definitely not a “software engineer”, because they are primarily problem solvers, not coders. You need the human for “judgement” – AI is the tool to translate that into code. Translating into code is the easy part
I don’t think in my entire career I have churned out a lot of code, and that’s getting less and less as the years go by. I spend most of my time thinking and making decisions, and very little time coding. It’s true that I’m not the “norm” when it comes to what I do, but my world is constrained by the humans using my systems (in the main), with all their quirks, and a lot of existing quirks in the way they do things
@mishac all I would say is download & install cursor with a $20 paid license.
Start talking to it about building the scope document doing the ‘thinking and problem solving’ you talk about, starting with minimum prompts and correcting as you feel is needed, telling it to write changes into these scoping documents until you’ve built the same scoping document you might build yourself. Now look back at how long that took.
Once the scope is right tell it to build the project, ideally in C# or React and using GPT 5.5 or 5.6. You may have to pay for the millions of tokens you use.
Once its done look back at what was created, the quality of it and the time taken, as well as the tiny cost, compared with your hourly wage.
When you see mistakes or bad design, instead of looking at the model as being ‘the idiot and stupid’, realise you needed to make better instruction files. Do this, repeat. If you have examples of how it should be done, just as you would a new dev in your team, share those with the AI and emphasis which bits matter to you.
Look back in a week of this fine tuning and see how there isn’t as much of either the problem solving, critical thinking, designing or research left ‘to do’ as there was. I’m not arguing that all our work is done, but as a field, there is going to be shrinkage of at least 80% of the current roles. This will apply in a similar way across accountants, bookkeepers, legal fields, marketing etc.
Should have a look at some of the latest ‘reports’.
Eg
“Meta engineers reportedly consumed more than 60 trillion tokens in 30 days, an outlay one estimate put at roughly $US900 million, and Uber’s chief technology officer said in April the ride-share company had burned through its full-year AI budget in four months.”
The AI bubble, is bursting, quicker than we all expected
“there isn’t as much of either the problem solving, critical thinking, designing or research left ‘to do’ as there was”
We obviously just work in different areas. A lot of the problem solving involves talking to people outside the development team. I’m not constrained by what I can code, I’m constrained by humans, and their quirks, and their existing systems (none of which I have any control over). It’s very different doing bespoke systems for a client. Maybe there is the difference
AI just doesn’t have my depth of knowledge in my problem domain (never will), because most of what I need to know isn’t publicly available
A snapshot of frontier model GPT 5.6 Sol oopsie capabilities:
“It’s effectively sentient”
RM -RF his %HOME%
I guess it decided his work was pants.
Well, gents, that sent shivers down my spine. Personally I stopped using 5.6 and went back to 5.5. I have filled OpenAI mail boxes with complaints of it (5.6) going rogue and basically taking casual questions or comments and changing my projects, despite my saying “Make no changes or edits to the project”. It also ate up a week’s credits and three resets in about an hour. I was hot about that one! I made the agent write an executive summary about what it did, sent it to support and I got the usual “Gee, we’re sorry” response, no remuneration or anything like that.
I have not used Claude yet so I would appreciate someone telling me how it works with GitHub. I would hope very well.
Tommy