>>Deepseek will be sanctioned and therefore no provider will offer it anymore.
That is possible inside the US. How do you do it all over the world? You have to convince every country in the world not use frontier models? Even worse how do you convince all the countries to not build their own models?
>>So how these companies and people manage to use these absurd amount of tokens is a mystery to me.
Absolutely!
I know some colleagues who are routinely spending thousands of dollars worth of tokens, I can't see to even max out the subscription limits even if Im working all the time. Curiously enough their output is lower too.
>>My impression is that the saying "Perl makes easy things easy and hard things possible" really does hold up.
People who haven't used Perl to its full power have little how idea just how magical a language it was/is.
Have seen people's jaws on the floor watch Perl guys do automation they always thought was impossible, even more so delivered in such record times. CPAN itself as an idea was way ahead of its time.
Around COVID I had a project delivered, in a week. I basically wrote a Perl script that wrote Apache Pig scripts(Basically using Perl like a macro facility, and Perl is great at anything text). It was a massive project which otherwise would have taken more than a year to deliver. When I did show the team what I had done, had them in total awe in the same way people look at Claude Code today. Nevertheless it was the same reaction, when the project got done, they were not comfortable that some programmers could do stuff like this, which seemed alien to the remaining.
I have a belief that proliferation of Java/Python stunted the growth of web dev, and a lot other industries. In some way the last decade was entirely slowed down by adoption of these technologies. If Ruby/Perl were there things would have been way better.
I do love what LLMs are doing to Java and Python today.
Software development was never supposed to be as slow as what Python and Java made it.
Have literally seen a 'architect astronaut' crap on Perl in some 2 hour long call, and eventually had his way with getting the management to approve Java. This was mid-2000s
When our team initially budgeted it, 4 guys over 6 months were enough to get this over the finishing line. The java team took over took more than 3 years, and close to 30 people. It was a AbstractClassFactoryFactorySingletonDispatcher mess with spring decorators all over. Which quite honestly was quite ironic because the original case against Perl was it was hard to read.
The java code was easily 30x more verbose, no body at the end knew how to maintain it. It was all about the guy getting to own his own team, promotions, bonuses, raises etc.
Have seen the same story repeat over and over again, everyone knew they wanted Python because they could get to inflate their headcounts.
Its one of things about tech, its not the good tech that wins, its the tech that helps with office politics wins at the end of they day.
After golang came along a lot of these java things too went out of fashion. Curiously enough golang does feel a lot perly to use. And Python has long moved away from its minimalism activism days. To that Python has transformed into the same feature bloat it once accused Perl of.
>>Turns out that Larry (and the team) were much better at language design than project management.
It is true many times to deliver quality products you can't have deadlines. But without a deadline you are never finishing a thing.
Unfortunately for Perl, Larry Wall, and several of its project leads(Patrick Michaud, Audrey Tang) at various times had major health issues. Time moves on, and people have to at times resign entirely from projects due to shifting priorities and personal problems. Parrot VM I guess went through a similar arc.
Other people have moved mountains to get Perl going. But with time people's priorities have entirely moved on. At one time, all Python programmers would do is bad mouth Perl all over the internet, and that never really stopped. Any body who saw a Perl programmer do over a weekend, what they would take a year to do in their language(especially Java and Python)- had a deep rooted seething envy at Perl and Perl programmers. So they went around almost on religious crusade to have Perl gone. This was done entirely to crush competition. They just didn't want other people to wield a power they didn't have. Lisp has had a similar arc of development over the decades.
Perl 5 development being entirely stopped for years further complicated this issue. Eventually as most of the Perl code in many companies bit rotted and died, newer projects were started in Python/Java. And of course Frontend stack entirely moved away to Node/React. We had mobile development of which Perl never was ever a part of.
By the time ML/AI era came into being Python was defacto the language of programming for these kind of tasks.
The best part is now in the LLM era, the whole idea of a programming language itself is pointless.
>>We used to romanticize stories of people born into lower socioeconomic conditions (say in the third world) and rising out of it through their intellect.
Many people think the social mobility came through intellectual feats, but when you look at it carefully, it came through agency and ownership.
Validated data coming out of LLM(Itself generated from a recursive/loop process, used to incrementally arrive at solutions) being using to improve the very LLM is a very powerful loop. And there is no real upper limit to this, at least not in the near future.
Like most exponential processes, the start is slow, but it gets fast very rapidly.
This is also what Anthropic has said, the ones training these models today(i.e 2025 - 2027) will be impossible to catch up with, let alone beat.
Mostly because we are looping AI to fix problems, and then the same data is used to improve AI. There is no upper limit to this.
Taken to its logical conclusion, this process needs a hardware scale that might even look laughably huge at this point. Its fairly obvious space is going to play a big role in the coming times.
I could be wrong, and I humbly accept it when Im proven wrong. But it does feel like a lot of people in top places know we are going to need all the energy and resources space has to offer to run this runaway intelligence.
Agency is just to keep moving, like stop at nothing and keep moving, no matter what. Movement generates information, that can be used to make further decisions, even if you actually don't hit your goal at bulls eye accuracy you do end up getting a lot far and learn a lot.
So we are kind of there at runaway intelligence today.
>>We just don't know what the maximum capability of AI is
For all theory purposes there is no limit. Thats what the latest loop engineering trend is about, you are asking AI to find solutions to a problem going by listing steps, and if solution not found in those steps, to treat each step as a separate problem and repeat the process until the master solution to the master problem is found.
Once a solution is found, or new data/insights are generated through this process, the LLM can be trained on this. So in theory you can just keep going like this forever.
Secondly. This is as close to agency you can build inside a machine.
Practically speaking, hardware is a limit. But that can scale up with time.
So we are already looking at some kind of runaway intelligence even if not sentient.
>>We don’t really need to bother using GPL/Apache licensed software because we can one-shot something of our own and not bother with giving back contributions.
Thinking of a open weight/source AI as gcc/perl was in the 1990s is more helpful line of approach to take here.
That's what the Fable harness felt like. You give it a goal and it could try to get there through the shortest path given the tree of possibilities to get there. Iteratively, or recursively.
Perhaps if we make a open coding AI, the design must be along these lines. Something that's easy to train, and serve from local machines. Albeit has loop / recursive hill climbing facilities built it. That way the model gradually keeps moving towards the solutions, in iterations/recursions.
Once this is done, other multi modal things could be pursued.
What sort of hardware are you using to run local models? And how do you use them?