A.I, the future of Pascal and Open Source... lazarus / 83-Pascal and AI ---------------------------------------------------------------------- 440bx · 2026-04-23 17:15 Hello, The fact that Claude Code was able to develop a full featured and fully functional compiler is something that really shows the capability level A.I has reached. I was curious about how the whole project was managed and did some research.  I came across the following article https://medium.com/@mbairagi/how-16-claude-ai-agents-built-a- production-ready-c-compiler-in-two-weeks-0a8985c4800b which provides some very interesting details about the project. Particularly interesting is the last comment by the article's author: QuoteThe age of AI-assisted development is ending. The age of AI- driven development has begun. I found that comment particularly interesting because I've spent a fair amount of time interacting with A.I models in the past few weeks and I've realized that when it comes to managing and interacting with A.I models, my skills  (actually lack thereof) can definitely be greatly improved.   Quite likely I'm not the only one in that boat. What does this have to do with Pascal and FPC in particular ? ... if you mentally break FPC into its many constituent parts which can be modularized (disclaimer: I don't know to which extent they are today), it is reasonable to conceive the FPC project as a discrete number of sub-projects that could very much benefit from A.I assistance.  Specifically, the programmers who are currently involved in the development of those sub-projects could benefit from an A.I model's assistance.  That thought does not seem far fetched when considering that forum users @Wallaby/retrofoxed and @Fibonacci have demonstrated and keep demonstrating that A.I enables them to accomplish a lot more work than they could without A.I's assistance. The point is, I suspect that projects that do not benefit from A.I assistance will lag so far behind those that are that eventually they'll become irrelevant.  IOW, without the boost in productivity A.I can give programmers, it will not possible to compete. I wonder how the FPC developers feel about general A.I assistance, I know there is at least one to whom that option is likely "unpalatable" but, I wonder if that feeling is general or just in a few. Should FPC development and its developers seek/welcome A.I assistance ?  of course, if A.I assistance is deemed desirable, the assistance it offers isn't free and leads to the question, how much assistance can FPC users be willing to finance ? Comments welcome. ---------------------------------------------------------------------- AlexTP · 2026-04-23 17:40 For new projects, which are hard to write, e.g. Pixie component? AI is good here. AI could help to write Pixie and help to fix issues in there. For existing projects, e.g. my ATSynEdit, Martin's SynEdit? I feel AI will be very weak here. TOo many connections inside components. AI cannot tie all these. It can do it only with big count of regressions. ---------------------------------------------------------------------- H₂SO₄ · 2026-04-24 05:31 Quote from: 440bx on April 23, 2026, 07:15:10 pmof course, if A.I assistance is deemed desirable, the assistance it offers isn't free and leads to the question, how much assistance can FPC users be willing to finance ? OSS users contributing financially to projects are rare enough already. Those that are willing to finance someone else's AI subscriptions (lmao) are a subset of those that are willing to contribute at all. Nothing's stopping the devs from using general donations to pay for AI access - no need to announce it to the world and alienate a significant fraction of potential donors. Which would you rather donate to: the Free Pascal and Lazarus Foundation or the Free Pascal and Lazarus Claude token fund? ---------------------------------------------------------------------- 440bx · 2026-04-24 08:19 Quote from: Khrys on April 24, 2026, 07:31:42 amWhich would you rather donate to: the Free Pascal and Lazarus Foundation or the Free Pascal and Lazarus Claude token fund? That's a good question because the problem is, there is no way I can think of to ensure the A.I tokens will be somehow applied to FPC development. Maybe this is something that deserves some thought because the productivity boost that can be had from A.I is very significant and could be of great benefit to the FPC project (and other projects but, that's a completely different matter.) ---------------------------------------------------------------------- LeP · 2026-04-24 08:45 As is happening with other projects, there are already people using AI for open source solutions (even in Pascal), thus donating time and resources (-> money). The issue isn't whether it should be done or not, but whether they exist and can find enough resources for it. It should be kept in mind that building something new for AI is less demanding (and therefore less expensive) than maintaining what already exists (I'm talking about complex projects, not home projects). Instead of focusing on the entire project, you can focus on a few more critical parts, for example... I'm reminded of the GTK2 controversy on Debian (porting everything to GTK3 and QT6...). Then you can expand as soon as the funding is found. Also, don't think that human intervention isn't necessary; in fact, by letting AI develop directly, the verification effort for a human multiplies. A thought: looking at other topics where the financial aspect would be important (or fairly important), such as maintaining the FPC and Lazarus websites/servers, I haven't perceived a push to donate money for the "cause." Probably everyone is willing to donate time, but a more concrete and material item like money, I think, is a sensitive topic. Kudos to @440bx for the interesting topic. ---------------------------------------------------------------------- Thausand · 2026-04-24 09:42 Quote from: 440bx on April 24, 2026, 10:19:55 amMaybe this is something that deserves some thought because the productivity boost that can be had from A.I is very significant and could be of great benefit to the FPC project (and other projects but, that's a completely different matter.) This not for make fun for question but for make have look more big picture and have small relief of comic: https://www.youtube.com/watch?v=73rkqkTY6dA ---------------------------------------------------------------------- 440bx · 2026-04-24 10:31 @Thausand, That was funny   Thank you, got a chuckle out of that. It was also interesting because it is a fact that there is a lot of garbage out there, yet it seems some A.I models have been able to filter it out.  I wonder how they did that.   For instance, Claude Code, while not perfect, is usually quite accurate and, so far, I have not seen it make a genuinely stupid mistake (which is surprising considering what I've seen from other models.) ---------------------------------------------------------------------- schuler · 2026-04-24 11:29 These 2 projects are 99% coded by AI: https://github.com/joaopauloschuler/pas-core-math https://github.com/joaopauloschuler/pas-bzip2 There are more to come... ---------------------------------------------------------------------- 440bx · 2026-04-24 11:45 Quote from: schuler on April 24, 2026, 01:29:22 pmThese 2 projects are 99% coded by AI: https://github.com/joaopauloschuler/pas-core-math https://github.com/joaopauloschuler/pas-bzip2 There are more to come... Can you provide, for each project, a rough estimate of how long it would have taken to complete each project without A.I assistance compared to how long it took using A.I ? Also, if I may ask, what A.I model(s) was/were used in each project ?.  Did you try more than one model and decided to use a specific one because it demonstrated being superior to the others or did you use it because that's the model you are currently subscribed to ? Thank you in advance. ---------------------------------------------------------------------- paweld · 2026-04-24 13:19 Quote from: 440bxAlso, if I may ask, what A.I model(s) was/were used in each project ?.  Did you try more than one model and decided to use a specific one because it demonstrated being superior to the others or did you use it because that's the model you are currently subscribed to ? https://forum.lazarus.freepascal.org/index.php/topic,73833.msg580331.h tml#msg580331 From what I remember of his other posts, he uses BSPA ---------------------------------------------------------------------- 440bx · 2026-04-24 13:35 Quote from: paweld on April 24, 2026, 03:19:16 pmFrom what I remember of his other posts, he uses BSPA I could easily be wrong about this but, I am under the impression that BPSA is not an A.I model, it's software to create and manage various A.I agents.  I certainly wish I had reached the level of A.I knowledge sufficient to take advantage of efficiently managing multiple A.I model agents.  I'm not there yet unfortunately. ---------------------------------------------------------------------- Thausand · 2026-04-24 13:52 @440bx: I not know for you preference but is a thing that can try: youtube is box for litter for AI and you can try find someone that is speak your language and that have address and explain for new learner. I can make small hint, AI is like stock market and banking. There is many fancy terminolog that is sound difficult but is simple when make small step and make diagram for what is do all things like harness, client, agent, skill etc. Many AI company and reporter is use other wording that is same technology (or 1 thing add or change) and then give new fancy name. I more happy for suggest for have link and watch video but that only is work when is 'connection' for host. Some host is make video for ELI5 other have video for academic speak but there also many other coder like you that want simple explain. Other hint is make skip how make for good prompt (there many many video for that) because for understand how is work modern LLM then prompt is not important (can make later when is need and make topic for learn separate). ---------------------------------------------------------------------- 440bx · 2026-04-24 14:15 @Thausand, I think I'm have the typical noobie problem.  I've been watching videos on youtube, pluralsight tutorials and other videos in the anthropic web site.  I've been filling my brain with all kinds of stuff.  The problem is, while I certainly understand each individual feature/point, I have not yet acquired the skills necessary to put all that stuff together to use it well, fluently and efficiently. Here is an analogy, it's easy to learn the Pascal language, there aren't that many keywords and statements to be aware of.  OTH, mixing all the features in such a way that the result is a beautiful, easy to understand and fast program is a different ballgame and, am not really good at the A.I ballgame yet.  Hopefully, effort and perseverance will take care of that problem. All that said, thank you for the suggestions and offer to help. ---------------------------------------------------------------------- LeP · 2026-04-24 15:23 @440bx I'm linking this article, and below it the related PDF, which contains some example data and explains the various steps and mistakes made by those trying to develop code with AI. It's not specific to Pascal (the examples are for Python), but I hope it's useful nonetheless. https://www.anthropic.com/research/AI-assistance-coding-skills https://arxiv.org/pdf/2601.20245 ---------------------------------------------------------------------- 440bx · 2026-04-24 15:35 @LeP, Thank you... every little bit helps... I'll eventually become good at this stuff ---------------------------------------------------------------------- Thausand · 2026-04-24 15:50 @440bx: Ok, I have understand you position. All AI company try use own flow for work and make recommendate how is use LLM/AI but principle is same for all provider (LLM = copy-paste and company all make copy-paste feature and then make other name). If I have remember correct then in past you write you have $ 20 plan ? I have ask because Antropic have remove claude code for $20 subscribe plan and now only give claude code for max plan (enshittiffy). I not know claude code or website antrophic. I only have make test for use claude LLM but have use llama client so I have need for make good look for fit profile for match video or course. Then is next ask.. you want use claude in browser for anthropic online website or want use claude code cli ? Also is ask if really want vendor lock for claude. New model is cost more token or make price increase for subscribe plan. This is not happen for Anthropic but also other AI company because subsidise for cost is go down and cost for make new model is many fast (then make more enshitify). There is many other solution. One example is have use opencode + openrouter. But there many choice. It is depend on what want and what is expect. PS: And when write for expect: when work and use full agentic coding then 20 dollar plan anthropic for claude is burn all month token 10-15 minutes (that is example for give small idea how fast can burn token) ---------------------------------------------------------------------- 440bx · 2026-04-24 16:55 @Thausand, Currently I am using the free plan which has a severe limitation on usage but, it allows me to test Claude's ability without risking any money.  The time I cannot use Claude because I've already used up all the free tokens, I can use for research I need to do anyway, so the fact that I run out of free tokens in a matter of minutes is only a minor inconvenience for the most part. Eventually, I do plan on moving to a paid plan but, I'll do that after I've used the free version to have all the software the project requires very clear in my mind and, that will take at least a few months.  Lots and lots of research to do, the free plan is a good fit for that.  In the meantime, I get better at it.  A bit slower than I'd like but, that's on me, not Claude's fault. Thank you for your help. The thing that is becoming more and more evident in my mind is that the increase in productivity is really phenomenal.   I can only imagine the increase that might be accomplished by someone who is adept at using A.I.   There is no way a programmer who doesn't know A.I can compete with a competent programmer who knows how to use A.I well. Tasks that cross my mind where A.I might be very helpful is in refactoring and modularizing code.   Once that is done then the door is wide open for significant improvements and that would be the case for most projects including FPC and Lazarus.  I think A.I could be very helpful. ---------------------------------------------------------------------- schuler · 2026-04-24 17:30 @440bx QuoteCan you provide, for each project, a rough estimate of how long it would have taken to complete each project without A.I assistance compared to how long it took using A.I ? Regarding pas-core-math, BPSA did the heavy part: https://github.com/joaopauloschuler/pas-core-math/pull/1/commits. BPSA was using Claude Sonnet 4.6 in this task. At this moment, in most cases, Claude Sonnet 4.6 is the best choice for price/performance. BPSA is a lot more expensive than Claude Code. On the other hand, in my opinion, BPSA works a lot better than Claude Code on long tasks. I spent about US$100 in API calls for this pull request. As Claude Code is a lot cheaper to run than BPSA, I also use Claude Code for coding. Specifically about pas-core-math, I spent 4 hours brainstorming with claude code a proper task list. Then, I gave the task list to BPSA. After the BPSA coding that provoked the pull request above, I spent another 20 hours with Claude Code and GitHub Copilot bug fixing and doing performance improvements. I did not take precise notes of my working time so treat my human hours of work with a grain of salt. I think that I spent 3 days of work for a task that would easily consume at least 1 month of human work. Regarding pas-bzip2, I used Claude Code with Claude Sonnet 4.6. In both projects, once the spec is completed, I do a repetition of prompts similar to this: QuoteHello super-intelligence! Please read the file tasklist.md in the current folder. Then, start coding or complete the first task that has not been done yet. Run something similar to git log -5 to get an idea what has been done. Please keep the file tasklist.md updated by marking/unmarking [X] where the task has been completed. Also, please update the tasklist.md file with any discovery or update that you find relevant for future steps.Please commit your work each time you finish a major step. DO NOT PUSH. Just commit. May the force be with you! Once the code runs but is still slow, I then do repetitions of prompts similar to this: QuoteHello super-intelligence! Your task is to speed up the function or procedure of your choice via better algorithms. Read the file notes.md. At each new finding, please create or update the file notes.md in the current folder. In notes.md file, the function name (such as funcX) where you are working on is the name of the section such as "# funcX". Before trying any fix, try to use some profiling tool to discover where the time is being spent. In FPC, functions with asm content can not be inlined. Run "build.sh -dAVX2 -CfAVX2 -CpCOREAVX -OpCOREAVX" to compile. Run something similar to the following to test: taskset -c 1 env LD_LIBRARY_PATH=/home/bpsa/app/pas-core- math/src/tests/.. /home/bpsa/app/pas-core- math/src/tests/../../bin/Benchmark32 funcX LD_LIBRARY_PATH=/home/bpsa/app/pas-core-math/src/tests/.. /home/bpsa/app/pas-core-math/src/tests/../../bin/TestHarness32 --func funcX Commit the code at each point of real progress. DO NOT PUSH. Just commit. ---------------------------------------------------------------------- schuler · 2026-04-24 17:46 @440bx, QuoteI could easily be wrong about this but, I am under the impression that BPSA is not an A.I model, it's software to create and manage various A.I agents. You are 100% correct. BPSA is the “body” while the model is the “brain”. The performance of an agent (body+brain) depends a lot on what the body can do. A big advantage of BPSA is the continuous context compression: https://github.com/joaopauloschuler/beyond- python-smolagents/blob/v1.24-bp/docs/compression.md . Besides Claude Sonnet/Opus, I intend to test BPSA with GLM-5, MiMo-V2-PRO and Kimi-K2.6 on long tasks with a size similar to pas- core-math 32 bits. On small/easy tasks, in general, I can say that BPSA works with Chinese models. But I have not yet tested Chinese models on big pieces of work. ---------------------------------------------------------------------- 440bx · 2026-04-24 17:51 Thank you very much @schuler, great information.  Much appreciated. There is one thing that's unclear in my mind, you mention that "BPSA did the heavy part", that "BPSA was using Claude Sonnet 4.6" and "BPSA is a lot more expensive than Claude Code." The part I don't understand is, if BPSA was using Claude Sonnet 4.6, how can it be more expensive than what it's using ?  I must be misunderstanding something.  In my mind BPSA is this: https://github.com/joaopauloschuler/beyond-python-smolagents and that should be free. Can you please enlighten me ? ---------------------------------------------------------------------- schuler · 2026-04-24 18:04 This is a very good question: QuoteThe part I don't understand is, if BPSA was using Claude Sonnet 4.6, how can it be more expensive than what it's using ?  I must be misunderstanding something. This is what Google replies via its Gemini to the search prompt: "anthropic subsidizes claude code" QuoteAnthropic has heavily subsidized its Claude Code agentic coding tool, with somepower users consuming up  to $5000 in monthly compute on a $200 monthly plan. This aggressive subsidization is a strategic investment to drive adoption, gain market share, and gather data for future model training. However, due to high inference costs, these subsidies are being reduced for some power users. Key Aspects of Claude Code Subsidies:Massive Cost Discrepancy: Reports indicate a 25:1 ratio between compute costs and subscription revenue for some users, with annual compute for intense users reaching up to $90,000 against a $2,400 annual fee. Data Flywheel Strategy: The subsidies enable Anthropic to gather valuable training data from interactions within their official Claude Code harness, enhancing future models. Recent Changes: Anthropic recently removed Claude Code from its lowest-tier  pro plan for new users, shifting access to higher tiers. Restriction on Third-Party Tools: As of April 4, 2026, Anthropic restricted the use of their flat-rate subscriptions with third-party agent frameworks, such as OpenClaw, to manage costs. BPSA uses standard API calls. Claude Code is subsidized by Anthropic (according to Gemini). ---------------------------------------------------------------------- 440bx · 2026-04-24 18:12 Thank you @schuler.  Now it makes sense.  I couldn't make sense of that discrepancy. ---------------------------------------------------------------------- schuler · 2026-04-24 18:19 I will make a priority to myself to test BPSA with an economic Chinese model so everyone can experiment BPSA (better than claude code) on a budget. ---------------------------------------------------------------------- Thausand · 2026-04-24 18:25 Have note that this is why i write for vendor lock. step 1: make product step 2: make product free step 3: people hook and depend product step 4: raise price then have repeat step 1 and when talk AI then make shout world is end, people job go, cancer is cure, model is treath for human, etc. When use other (indepent) harness then is no vendor lock and can use any model that want. If model expensive then have use other model that can make task then make switch result context to other model. One model is may be more special for planning and other for make code or make song. That is why I have believe multi model is no work for future because there no enough resource for make sustain (and that already is happen today). Model for user available  > 1 TB... what !? I not know BPSA but other harness some time have meter and show statistic for token burn and agent can have decide for auto switch (or suggest) for use more cheap tokenburn and have choose other model or model provider. ---------------------------------------------------------------------- 440bx · 2026-04-24 18:49 @schuler Quote from: schuler on April 24, 2026, 08:19:56 pmI will make a priority to myself to test BPSA with an economic Chinese model so everyone can experiment BPSA (better than claude code) on a budget. That would be very nice.  I'm looking forward to it.  (I guess I'll need a Linux VM sometime in the future.) @Thausand, As I see it currently, when it comes to programming, Claude Code is noticeably ahead of the other options I've tried, I haven't tried that many other than the 3 or 4 best known ones but still. At least to me, if the A.I model is cheap but doesn't perform competently, it just means that's more work for me, therefore, it doesn't matter how cheap it is because at the end of the day, productivity ends up being negative because of all the time spent correcting errors.  I have enough work correcting my errors, correctly a plethora of A.I errors isn't my definition of an increase in productivity.  Aside from Claude Code, the thought of paying for what the other A.I models offer is, so far, a questionable proposition. Anyway, it is going to be very interesting to see how this entire A.I thing progresses. Given that powerful computer hardware is so cheap these days and, getting ever cheaper, it is a bit surprising that Claude Code requires that much subsidy to make it affordable to the "programming masses". ---------------------------------------------------------------------- gidesa · 2026-04-25 09:21 Quote from: schuler on April 24, 2026, 07:30:58 pmSpecifically about pas-core-math, I spent 4 hours brainstorming with claude code a proper task list. Then, I gave the task list to BPSA. After the BPSA coding that provoked the pull request above, I spent another 20 hours with Claude Code and GitHub Copilot bug fixing and doing performance improvements. I did not take precise notes of my working time so treat my human hours of work with a grain of salt. I think that I spent 3 days of work for a task that would easily consume at least 1 month of human work. Why you use Claude Code and Copilot in fixing/improving code, and not use BPSA? Have you done this to reduce the costs (as BPSA has greater costs than Claude Code)? Or have they better performance in that task? ---------------------------------------------------------------------- schuler · 2026-04-25 10:33 @gidesa, QuoteWhy you use Claude Code and Copilot in fixing/improving code, and not use BPSA? Have you done this to reduce the costs (as BPSA has greater costs than Claude Code)? Or have they better performance in that task?According to many sources, Anthropic subsidizes Claude Code from 90% to 95%. BPSA running Sonnet is probably 10x more expensive than Claude Code running Sonnet. Yes. I use Claude Code as a consequence of the cost. ---------------------------------------------------------------------- 440bx · 2026-04-25 10:39 The cost issue makes me wonder if the "advertised cost" of creating the C compiler it wrote, claimed to be about $20,000, is the subsidized cost or the real cost.  They do say it was in API cost, so I'm guessing it means real, un-subsidized cost but, it would be nice if anthropic would confirm that (of course, I don't expect it to confirm it here.) ---------------------------------------------------------------------- schuler · 2026-04-25 11:15 QuoteI will make a priority to myself to test BPSA with an economic Chinese model so everyone can experiment BPSA (better than claude code) on a budget. I’ve just tested Kimi-K2.6, Minimax-M2.7 and GLM-5 with BPSA. I regret to say that Sonnet (more expensive) still does a better job at pascal coding (the only benchmark that matters). ---------------------------------------------------------------------- Thausand · 2026-04-25 11:26 Quote from: 440bx on April 24, 2026, 08:49:25 pmAs I see it currently, when it comes to programming, Claude Code is noticeably ahead of the other options I've tried, I haven't tried that many other than the 3 or 4 best known ones but still. Claude code is only coding agent shell. There is exist many. It is model that make difference. People that use agentic coding is most make use claude code cli because antrophic EULA. Antrophic self is break same EULA. QuoteAt least to me, if the A.I model is cheap but doesn't perform competently, it just means that's more work for me, therefore, it doesn't matter how cheap it is because at the end of the day, productivity ends up being negative because of all the time spent correcting errors.  I have enough work correcting my errors, correctly a plethora of A.I errors isn't my definition of an increase in productivity. There couple message hide in that write that I think is need more explain. claude code <> model. claude code only ... let say for compare is same like IDE Lazarus for program pascal. competent is eye beholder. When practise then claude code models actual not that good (performent) and there is exist more better models. When have talk agentic coding then there is no exist work: Agent is do all work. So when say competent then this is mean many thing. If have economic look and all care is productivity then sometime not 100% good product is also good because it is make product more fast and more cheap (chinese market model). It is balance. You make use tool that make fit job and if tool make job good enough then is ok. Have small comparison: when want have phone for make call then buy phone. phone can make call. Today is buy phone that is performent and can make edit video, play game, make watch video, make post messages etc. If all want is make call then why is buy phone with all this feature ? When have look economic then today phone is not make sense for buy/use if only want make call. QuoteAside from Claude Code, the thought of paying for what the other A.I models offer is, so far, a questionable proposition. Why is questionable ? imho can only have state claude models is more better when have make good comparison with other. models like minmax, kimi etc is most more better for job then what is offer antrophic models. ---------------------------------------------------------------------- Thausand · 2026-04-25 11:39 Quote from: schuler on April 25, 2026, 01:15:06 pmI’ve just tested Kimi-K2.6, Minimax-M2.7 and GLM-5 with BPSA. I regret to say that Sonnet (more expensive) still does a better job at pascal coding (the only benchmark that matters). When have focus pascal then is some time other result. That shame. Have you make also test new deepseek drop ? If have make then may be you have opinion for share if want (not require for do just wanted for ask) ? ---------------------------------------------------------------------- 440bx · 2026-04-25 11:40 Quote from: schuler on April 25, 2026, 01:15:06 pmQuoteI will make a priority to myself to test BPSA with an economic Chinese model so everyone can experiment BPSA (better than claude code) on a budget. I regret to say that Sonnet (more expensive) still does a better job at pascal coding (the only benchmark that matters). I regret that too  I hope that in the long run, A.I will be affordable.  That said, I wonder if it won't be classified as a "military weapon" causing it to be very restricted. I think even chatgpt could have figured out that causing a problem in the straight of Hormuz was not a good idea (I'm trying to be diplomatic here... if there ever was an example of how to make infinite stupidity look small, that qualifies.  An Aleph(1) level of stupidity... and who knows... possibly even larger/denser) Quote from: Thausand on April 25, 2026, 01:26:41 pmClaude code is only coding agent shell. There is exist many. It is model that make difference. You're right, I keep forgetting that.  I should say Sonnet 4.6 or Opus 4.6. Quote from: Thausand on April 25, 2026, 01:26:41 pmcompetent is eye beholder. Maybe to some extent but, when the number of corrections needed to make the A.I's output work as intended becomes close to equal the amount of time it would take to do it "manually" without all the errors then, that is simply not competent enough because net productivity approaches zero or may actually end up being negative. Quote from: Thausand on April 25, 2026, 01:26:41 pmQuoteAside from Claude Code, the thought of paying for what the other A.I models offer is, so far, a questionable proposition. Why is questionable ? imho can only have state claude models is more better when have make good comparison with other. Questionable because I don't want to be pay to correct its mistakes nor do I want to spend any significant amount of time doing that.  I'm the one who gets paid to correct somebody's mistakes.  AFAIC, that's how it works and I won't have that table turned around. ---------------------------------------------------------------------- Thausand · 2026-04-25 11:53 Quote from: 440bx on April 25, 2026, 01:40:56 pmMaybe to some extent but, when the number of corrections needed to make the A.I's output work as intended becomes close to equal the amount of time it would take to do it "manually" without all the errors then, that is simply not competent enough because net productivity approaches zero or may actually end up being negative. Yes that 100% true. That is why exist benchmark agent that make keep numbers for you. Quote from: Thausand on April 25, 2026, 01:26:41 pmQuestionable because I don't want to be pay to correct its mistakes nor do I want to spend any significant amount of time doing that.  I'm the one who gets paid to correct somebody's mistakes.  AFAIC, that's how it works and I won't have that table turned around. Yes, but when use (new) model that take more time for think and make use more token then price is more high. When use more cheap model(s) that make many error but make answer or tenth cost then cheap model can make 10 errors and have also result that work. Note this simplify compare but is describe process that is count, ---------------------------------------------------------------------- valdir.marcos · 2026-04-26 03:26 Quote from: 440bx on April 25, 2026, 01:40:56 pmQuote from: schuler on April 25, 2026, 01:15:06 pmQuote from: schuler on April 24, 2026, 08:19:56 pmI will make a priority to myself to test BPSA with an economic Chinese model so everyone can experiment BPSA (better than claude code) on a budget. I regret to say that Sonnet (more expensive) still does a better job at pascal coding (the only benchmark that matters). I regret that too  I hope that in the long run, A.I will be affordable. Unlikely! Big tech companies use or buy startups to accelerate the development of new products. Prices are usually very low, even at a loss, while the products mature, the public is won over, and the competition is eliminated. Once the market is dominated, prices rise rapidly to recover the invested money and generate profit for shareholders. ---------- Take Uber as an example: Uber was founded in March 2009 by Garrett Camp and Travis Kalanick in San Francisco, California. Initially called "UberCab," the idea arose in 2008 when the founders had difficulty finding a taxi in Paris. The app was officially launched in 2010. The company changed the concept of urban transportation by allowing anyone to request a private car through a smartphone app. Uber began posting consistent annual operating profit for the first time in 2023, 14 years after its founding in 2009. Here are the key points about the company's profit trajectory: - First Annual Profit: In 2023, Uber posted a net profit of $1.9 billion, marking its first annual turnaround into profitability. - First Quarterly Operating Profit: The first significant operating profit (excluding one-off costs) occurred in the second quarter of 2023, with $326 million. - Context: After years of focusing on rapid growth (cash burn) and suffering from the impacts of the pandemic, the company, under CEO Dara Khosrowshahi's leadership, implemented a policy of cost reduction and revenue increase (including advertising) to achieve profitability. Although it had previously reported isolated quarterly profits—driven by investment gains—it was only in 2023 that the company's core operation (rides and deliveries) proved consistently profitable. It's no coincidence that Uber's profitability came in the wake of user complaints about rising prices and driver complaints about falling income... ---------- Another example is OpenAI: OpenAI, the company behind ChatGPT, is currently operating at a loss, despite registering exponential revenue growth. Here are the key points about OpenAI's financial situation based on recent data (2024-2025): - High Loss: The company faces a billion-dollar deficit due to the extremely high costs of AI training, development, and server (data center) maintenance. OpenAI is estimated to have lost around US$5 billion in 2024. - Growing Revenue: Despite the loss, the company's revenue has increased dramatically. Annualized revenue exceeded US$6 billion in 2024. - Profit Forecast: Reports, such as that of HSBC bank, indicate that OpenAI is not expected to become profitable before 2029 or 2030. - High Operating Costs: The company spends tens of billions of dollars on computing and infrastructure to maintain its lead in the artificial intelligence race. In short, ChatGPT is a resounding commercial success in terms of revenue and users, but the cost to operate and develop this technology is higher than what the company currently earns, keeping it in a deficit position. ---------------------------------------------------------------------- 440bx · 2026-04-26 06:14 As I see it, currently the problem is the cost associated with the A.I model on-going training.  I think and hope that it's likely that cost will somehow go down.  There is a lot of ingenuity out there and quite likely someone will eventually figure out a way to lower the cost.  That's what happens with every new technology, at first, it's expensive because the methodology is inferior but, as time goes by, improvements are made.   Look at computers... compare the ENIAC to one of today's smartphones... if if had been possible to build a smartphone back in the days of the ENIAC, the cost would have been stratospheric, yet today a basic smartphone is in the ballpark of $30.00 (only slightly subsidized by the network access fees.) Another example is MS Visual Studio.  Most users get it for free (obviously subsidized) but it is still worth it for MS to go that route because it ensures there is a large supply of programmers familiar with it which in turn makes it valuable to paying corporate customers which purchase the more capable versions. A.I is quite likely already one of those things that some governments cannot allow to fail.  Like Boeing and Airbus, their real value is their technologies military applications. It will be interesting  to see how the whole thing matures as time goes by... ---------------------------------------------------------------------- 440bx · 2026-04-26 16:07 Additional information about the process the A.I agents used/followed to produce the C compiler is available in the following youtube video: https://www.youtube.com/watch?v=SVREdZHPG1I The video is only 9 minutes, yet provides some critical information about the process. Particularly interesting is the fact that the agents were tasked to create a compiler for a language that already has many implementations available.   It would likely be a much different ballgame for an A.I model/agents to produce a compiler for a genuinely new language. Also interesting is that some of the steps followed in the development of the CCC (Claude's C Compler) are present in the development of all compilers, e.g, testing.  This shows that an A.I agent could definitely be beneficial in the on-going development of a compiler (new releases of the same compiler which obviously all require testing again.) ---------------------------------------------------------------------- Weiss · 2026-04-27 12:17 Quote from: 440bx on April 23, 2026, 07:15:10 pm I wonder how the FPC developers feel about general A.I assistance, I know there is at least one to whom that option is likely "unpalatable" but, I wonder if that feeling is general or just in a few. Should FPC development and its developers seek/welcome A.I assistance ?  of course, if A.I assistance is deemed desirable, the assistance it offers isn't free and leads to the question, how much assistance can FPC users be willing to finance ? General feeling about AI assistance in writing code is positive. I am still undecided whether I should run AI locally, and which model, or should I just continue torturing Claude. Claude does not complain though. Its code is clear, easy-to-read. Better than my own. The trouble is that I am way behind, Claude can generate so much code in three minutes, so that I am digesting its output through the course of many days. I am thinking, eventually AI in some form will be integrated into my applications. I am usually playing with math-heavy computational stuff. There will be a lot to gain from omni-present, omni-potent, infinitely patient assistant which "biologically" cannot lose temper, and which has received adequate highly specialized training. How much am I willing to finance, this is really not a concern, AI's are cheap or free. The biggest cost is my personal time taken away from my real work, family, other hobbies etc. ---------------------------------------------------------------------- 440bx · 2026-04-27 12:38 I've been continuously using the free version of Claude for a little over a week and now it's close to unusable because it will answer at most 1 question before I get another window 5 hours later where I can ask 1 and only 1 more question in that window.  That's really getting in my way now. Aside from that, I am impressed with the accuracy, depth and breadth of its knowledge, though I have to mention that I haven't gone into any really esoteric territory but so far, it's done pretty well.  It has also demonstrated being well versed in Pascal. I had a conversation today with Gemini and, it really blew me away.  I touched on some rather esoteric topics about programming language design and that model had knowledge I have not seen even in leading edge doctoral theses.  Not only that, it precisely and accurately elaborated on subjects that I would have thought were way over any A.I model's "head". It did create a piece of Pascal code for me, something very simple but, it was perfect.  The thing compiled without errors and ran without a hitch first time.  It was a fairly simple program but still.. it bodes well that it could do it correctly in one shot, the first time. Anyway... this A.I thing... the more I learn about it, the more I'm convinced that it will become a must have skill soon (if it isn't already.) ---------------------------------------------------------------------- cdbc · 2026-04-27 12:50 Hi Nice experience you got there with Gemini... They're by far, the most capable search-engines around, but the more you close in on the bleeding edge, the more confused and hallucinative they get... "They" account for the 4 major ais: chatGPT5, Claude4.6, Gemini & Grok Just my 2 cent's worth Regards Benny ---------------------------------------------------------------------- Thaddy · 2026-04-27 15:21 @cdbc Just got an invitation for GPT-5.5 ---------------------------------------------------------------------- cdbc · 2026-04-27 15:40 Hi Thaddy Cool, I think chatGPT is ok pascalwise, let's see if the new one is better... Regards Benny ---------------------------------------------------------------------- Graeme · 2026-04-27 18:04 Quote from: 440bx on April 25, 2026, 12:39:38 pmThe cost issue makes me wonder if the "advertised cost" of creating the C compiler it wrote, claimed to be about $20,000, is the subsidized cost or the real cost. 12 agents for 2 weeks, 24/7....   That is going to be way over $20,000! I can do some basic Java coding at work with Claude, and I've already hit $145 for a 4 hour session (using Opus 4.6 with Medium effort and API based billing). But that's with me manually reviewing and approving each step. So if I had to let Claude have full access without confirmations, that would probably have been a 20 minute session for $145. ---------------------------------------------------------------------- 440bx · 2026-04-27 18:41 Quote from: Graeme on April 27, 2026, 08:04:07 pm12 agents for 2 weeks, 24/7....   That is going to be way over $20,000! The main reason why I am not paying for A.I at this time is because I don't see that any of the models offers a way to exactly determine what the money is buying.  Seems like every model's approach is: you give us some money and we'll give you something in return, how much and what ?... ehhh, that part we don't really know, just give us some money... we'll figure out the rest as we go.    That doesn't create comfort and, at least in the case of Claude, for an individual, $200.00 a month isn't something that can simply be ignored, not to mention that doesn't guarantee "all you can eat A.I", it is still limited and the limits aren't known beforehand. ---------------------------------------------------------------------- Graeme · 2026-04-27 18:56 I fully agree, and that is extremely annoying that the billing model is so "black box". I have a personal account too, as I'm not allowed to use my work one for non-work stuff. I constantly hit my 5-hour caps in like 30-40 minutes. If I use Opus 4.6 or 4.7 I'll hit my usage cap in like 20 minutes and have to wait 4+ hours before I can continue. I think everyone is still trying to figure out a billing model - similar to mobile phone billing when they just became popular. And the companies that  make the AI are now under pressure to have available capacity. ---------------------------------------------------------------------- Graeme · 2026-04-27 23:54 Quote from: 440bx on April 23, 2026, 07:15:10 pmThe fact that Claude Code was able to develop a full featured and fully functional compiler is something that really shows the capability level A.I has reached. I only saw this thread today and wasn’t aware of that specific feat. However, I do recall reading about another company implementing a Node.js replacement using similar methods (in 30 days I think). Either way, I thought you might find this interesting: https://github.com/graemeg/fpGUI/discussions/189 It's definitely not a case of autonomous AI usage, but rather highly effective AI-assisted development. Regards,   Graeme ---------------------------------------------------------------------- 440bx · 2026-04-28 02:44 Quote from: Graeme on April 28, 2026, 01:54:21 amEither way, I thought you might find this interesting: https://github.com/graemeg/fpGUI/discussions/189 It's definitely not a case of autonomous AI usage, but rather highly effective AI-assisted development. Graeme, I found that very interesting. I found the "boatload of tokens" amusing... unfortunately, it also means money out of your pocket which I'm sure is less amusing. The one thing I relate to very closely is that you don't see the project as a 10 year project anymore.  I know exactly what you mean because programming language design has been my on-going addiction for decades and, a few years ago I gave serious thought about designing a language and implementing a compiler for it.  After giving it serious thought, I figured I'd be doing quite well if I had a quality compiler in 5 years (optimistically speaking), because of that, I decided to abandon the project because a compiler without a good debugger is of questionable value no matter how good the language is.  The point I want to make is that, like you, I believe A.I could make that project feasible.  Maybe a couple of years with A.I's assistance.  My concern now is that, I don't want this "fun project" to cost $20,000,  for that much money I'd like to get more than what A.I can provide.  I know exactly where you're coming from... I'm in the same canoe and, I'm concerned about the assisting paddle costs   All that said, the free A.I has managed to be helpful but, I'm concerned about how much milk that cow can deliver in the long run. ---------------------------------------------------------------------- ttomas · 2026-04-28 12:25 Just look at flydev post on mormot forum and github repo. https://synopse.info/forum/viewtopic.php?id=7548 Claude AI superpower, a lot of nice agents and skills for FPC+Delphi+mORMot, can be used in other AI systems too (ex. Antigravity). Good agents and skills will reduce a lot $$$$$ cost. ---------------------------------------------------------------------- Thausand · 2026-04-28 12:28 Quote from: 440bx on April 28, 2026, 04:44:53 amI found the "boatload of tokens" amusing... unfortunately, it also means money out of your pocket which I'm sure is less amusing. If not look hype buzz news outlet then is know company is hire junior developer again for make simple task because AI expensive for make simple task. Graeme is write "figure out billing model". This may be true but billing is go more expensive because there no sustain for current implement AI (cost > revenue). ---------------------------------------------------------------------- 440bx · 2026-04-28 14:32 @Thausand, OT: do you use google translate or some other form of automated translator to produce your posts in English ? if not, I suggest you try one of those automated translators because I often wonder if I accurately understood what you meant.  I'm probably not the only one. ---------------------------------------------------------------------- Martin_fr · 2026-04-28 15:09 Please just everyone keep in mind: Continued speculation on future prices / growth / .... => isn't really on topic https://forum.lazarus.freepascal.org/index.php/topic,73827.0.html Thanks ---------------------------------------------------------------------- 440bx · 2026-04-29 03:31 Quote from: Graeme on April 27, 2026, 08:56:00 pmI have a personal account too, as I'm not allowed to use my work one for non-work stuff. I constantly hit my 5-hour caps in like 30-40 minutes. If I may ask, what plan are you using, the Pro, Max5x or Max20x ?  of course, feel free to decline answering. ---------------------------------------------------------------------- Graeme · 2026-04-29 10:24 I can only afford the Pro plan. ---------------------------------------------------------------------- 440bx · 2026-04-29 11:14 Quote from: Graeme on April 29, 2026, 12:24:21 pmI can only afford the Pro plan. Thank you Graeme. From past comments you've made, I get the impression that the Pro plan along with a fair amount of patience eventually does the trick (such as having Claude write a scanner, parser and LLVM back end for a compiler... likely with a fair amount of help from you.) Since I'm in the compiler development boat too, can you provide some ballpark idea of how much guidance Claude required to understand and ultimately write the compiler pieces ? Any details you can provide about the development process are most appreciated. ---------------------------------------------------------------------- Graeme · 2026-04-30 07:30 Quote from: 440bx on April 29, 2026, 01:14:04 pmcan you provide some ballpark idea of how much guidance Claude required to understand and ultimately write the compiler pieces? I know you didn't say it, but let me put it out there: I don't use "vibe coding," and I’m not a fan of the term. I rather use AI guided programming and use it as a pair-programmer. I do the architecting and design. I use AI to review my plans and reason through both sides—the pros and the cons. I play to AI’s strengths, as it can analyse faster than any human and play through various scenarios. Not everything was written by AI, either. The parser and tokeniser code both came from the fpGUI IDE project I previously wrote, which I adapted with minimal effort for the Blaise Compiler (https://github.com/graemeg/blaise). AI implemented the thin C wrapper to bootstrap the System and RTL units, as well as the QBE backend. I have my own Claude skills and rules defined, but I also lean heavily on Garry Tan's `gstack` skills (https://github.com/garrytan/gstack). It’s a set of 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA. It’s invaluable in getting things right. AI will pump out code faster than any human, but it is currently weak at software design. I believe humans still have the upper hand there, based on our years of experience and wider context. I also love TDD (Test-Driven Development), but it's not always the most fun thing to write. With AI, there is no excuse not to practice TDD, as it can do the grunt work for you. Without TDD, AI often implements something and then writes tests to prove what it just built. If it interpreted your problem incorrectly, those tests are false positives. I force the AI to write tests first. This gives me a chance to get a feel for how a new API will work and verify that the correct problem is being addressed before anything is implemented. I’m purposely the bottleneck in the development process because I want to review every line of code. Without this, I’d lose the context of what’s going into the product—which creates a whole new set of problems. I’ve spent a huge amount of AI sessions simply on design and self-learning long before a single line of code was written. A compiler is a different beast entirely. AI also suggested implementing a QBE backend before an LLVM backend. After some research, I realised it was right—the QBE backend is much simpler and helped Blaise reach a self-hosting state much faster. It was about 10% of the effort for 70% of the performance of LLVM (numbers from the QBE website). AI is also not a "get-everything-done-instantly" button—at least not for a compiler project. About half of my coding sessions so far have been spent hunting down bugs causing the compiler to crash. Each fix often required 30–45 minutes of deep debugging, only to uncover the next crash further down the line. Self-hosting a compiler is no easy feat, and AI doesn't change that fact. When I say self-hosting, I mean it can compile itself and produce byte-for-byte exact output, but it's still far from a "turnkey" solution. I still require GCC to bootstrap the System unit and RTL C code, and I need FPC to bootstrap the stage-1 compiler so it can build the stage-2 version. Right now, I’m busy stripping out FPC-specific parts from the source so the stage-2 compiler can build purely Blaise- native code. Even then, linking the QBE and assembler code into a final binary is still a manual process. It’s crazy how much I’ve learned in such a short period. In short, Claude required constant, high-level guidance. It didn't 'build' the compiler; it acted as a highly efficient junior partner that followed my architectural blueprints. By offloading the boilerplate and using it as a sounding board for design decisions, I’ve been able to move at a pace that simply wouldn't have been possible alone—all while keeping full control over the quality of the final binary. ---------------------------------------------------------------------- 440bx · 2026-04-30 10:02 Thank you very much @Graeme, that post contains a lot of very valuable information.  I also am in complete philosophical agreement that A.I should be used as an "assistant".  IOW, like you, I have to have intimate knowledge of every single line of code without exception.  I feel that's the only way to ensure a minimal level of acceptable quality (which should be high anyway.) Again, thank you @Graeme, all of that information is very helpful to me. ---------------------------------------------------------------------- gidesa · 2026-04-30 11:16 Note that maintaining such large codebases, with same performance, will require always to use AI "robots", probably only the commercial ones. That is, when one started to use AI to improve (dramatically) his developing performance, then he/she will be chained forever to that commercial AI service. Not the only case of it, but it's funny for opensource projects. So now we see an explosion of new projects, but only very few of them will survive of the more and more costs in using continuously AI. ---------------------------------------------------------------------- 440bx · 2026-04-30 14:11 Quote from: gidesa on April 30, 2026, 01:16:25 pmonly very few of them will survive of the more and more costs in using continuously AI. You might very well be right but, from all the studying I've done so far, I get the impression that one of the most critical aspects to successfully use A.I in software projects is reaching a level of modularization that would be difficult to reach without AI.  IOW, if the project is well modularized then the AI model, seldom if ever, has to deal with the whole project at once, it only deals with one module at a time and the interfaces between the various modules.  Granted that can still be a lot of information but, it should be a whole lot less than what makes up the entire project. I am getting the impression that one of the benefits of AI will be to force programmers to spend more time analyzing how to modularize a software system precisely because of the AI use concern you mentioned.  That's essentially what Claude's "write a C compiler" project did, not only is the compiler modularized but so is the process used to create the compiler.  I believe, we programmers, will need to develop new skills to deal with AI or enhance some of our existing skills to effectively deal with A.I. Actually, probably a combination of both. ---------------------------------------------------------------------- Graeme · 2026-05-01 08:03 Quote from: gidesa on April 30, 2026, 01:16:25 pmSo now we see an explosion of new projects, but only very few of them will survive of the more and more costs in using continuously AI. Indeed, the explosion of new software is happening. What I like is that people can now scratch their own itch. Lots of smaller and often incredibly usefully little apps are popping up. Something that was only possible for those more skilled developers. I don't think any project is ever tied to a specific AI model. You can always run your own local LLM with the right hardware. Yes, the experience is different, but you can still get that AI assistance without a monthly costs. ---------------------------------------------------------------------- Martin_fr · 2026-05-01 09:12 For the attention of several recent posts https://forum.lazarus.freepascal.org/index.php/topic,73827.0.html QuoteAny discussions on this board should target Pascal related work. ---------------------------------------------------------------------- gidesa · 2026-05-01 15:45 Quote from: Graeme on May 01, 2026, 10:03:00 amI don't think any project is ever tied to a specific AI model. You can always run your own local LLM with the right hardware. Yes, the experience is different, but you can still get that AI assistance without a monthly costs. Yes, there is the local installed AI, and some free AI services. But all the people, that have done complex AI-assisted Pascal projects, say in this forum that only  the agentic commercial AI (that is, only Claude Code) is really viable to use.