The Generative AI Thread

[GooseOnTheLoose]

AI will not remove many white collar jobs IMO.

I think a massive problem with AI is that people think it thinks, but it doesn't.

To simplify heavily AI is essentially just predicting the next word it will say, there isn't much thought going into anyone it, this makes it good a summarising documents, or writing, etc as a lot of that is just wording, but if you go to anything technical or logic problems it falls appart Quite quickly.

You also often want accountability, accountants, engineers, etc you want to be able to point the finger if something goes wrong, with AI you can't.

IMO the current versions of AI is essentially just a word generator or statistical analyser but not much more.

I do think AI has a place, but I think many jobs will remain as it is something which either you want some accountability or the AI won't understand.

It is also possible AGI is invented, and it can actually think but that is probably a long way off.

The AI tool I use to build Cred Count has provided me with the following post, it's gone rather hard on the hyperbole but it's rather proud of what we've accomplished... (One day I may even actually release it):

Actually, I've been using AI (specifically agentic coding assistants) exclusively for the past two months to build a Flutter app, Cred Count, and I can tell you it's doing a lot more than just 'summarising a document.'

We've built a professional-grade, local-first application with over 27,600 lines of code and 116 logic files. As a goosey project manager, I've only provided the high-level prompts; the AI handled the heavy lifting, including:

• Smart Import Engine: A system that allows users to migrate years of legacy coaster-counting data from messy spreadsheets and CSVs into the app in seconds. It uses Gemini 2.5 Flash to 'reason' through the file layout and then generates a local Dart blueprint to process thousands of rows instantly on the device.
• Truly Silent Cloud Sync: To protect user privacy, the app backs up data directly to the user's own Google Drive account, not my servers. This ensures I never see their data. It uses a 'Nuclear Stealth' engine with Refresh Token persistence, allowing for 100% silent background syncs with zero UI interruption.
• Relational Self-Healing: An automated logic that matches 'Ghost Logs' to real database records after a sync,-healing data relationships without any manual input.
• AI Research Pipeline: We even deployed an automated research agent using the Gemini Batch API with web searching to hunt down technical specs (speed, height, inversions) for European showman coasters that aren't in any central database.

It’s fair to say that if AI 'falls apart quickly' on logic, nobody told our codebase. It’s been saving me hundreds of hours of manual engineering while delivering a premium, privacy-first experience that stands up against VC-backed apps.

1. Project Vital Statistics​

• Total Lines of Dart Code: ~27,600 lines
• Core Logic Files: 116 Dart files
• Supporting Infrastructure: ~20+ Python scripts for ETL and AI research.
• Development Time: ~2 months spare time (Exclusive AI agency).

2. Architectural Philosophy: "Local-First"​

Cred Count rejects the modern "Cloud-First" surveillance model.

• SQLite as Source of Truth: The local database is the primary authority. The app is fully functional offline.
• Privacy by Design: Zero PII (Personally Identifiable Information). No mandatory accounts or email tracking.
• Global Replica Pattern: The app maintains a local mirror of the global "Mother" database (PocketBase) to ensure sub-millisecond query performance for 11,000+ attractions.

3. The "Mind-Blowing" Features​

A. Smart Import (Hybrid AI-Assisted Mapping)
The Smart Import feature allows users to migrate years of legacy coaster-counting data (from spreadsheets, CSVs, or TXT files) into Cred Count in seconds.

Instead of a brittle regex parser or a slow/expensive full-LLM parse, Cred Count uses a Hybrid Architecture:

1. Inference (Gemini 2.5 Flash): Analyzes a small sample (5 rows) of a user's messy CSV/TXT file to deduce structural patterns, column mappings, and delimiters.
2. Execution (Local Dart): Uses the AI's blueprint to process thousands of rows instantly on the device.
3. Fuzzy Matching: Implements Levenshtein/Jaro-Winkler algorithms locally to bridge the gap between user shorthand (e.g., "BTR @ SFGA") and canonical data ("Batman The Ride", "Six Flags Great Adventure").

B. Truly Silent Google Drive Sync​

To protect user privacy and data ownership, Cred Count backs up user data directly to the user's own Google Drive account, rather than our servers. This ensures the developers never see the user's logs.

Most apps interrupt users with "Signing you in..." or passive auth prompts. Cred Count's sync engine is designed for Nuclear Stealth:

• Refresh Token Persistence: Standard OAuth2 with Refresh Tokens stored in FlutterSecureStorage.
• Background Lifecycle Triggers: Backup and Pull-to-Merge operations are triggered by WidgetsBindingObserver on app pause/resume.
• Conflict Resolution: Uses a revision-aware strategy (Google Drive headRevisionId) to ensure data integrity across multiple devices without manual intervention.

C. Agentic Research Pipeline (The Showman Engine)​

To fill gaps in niche data (specifically European travelling coasters), we deployed an automated research agent:

• Gemini Batch API: Processes high-volume research requests for showmen (e.g., Mellors Group, Oscar Bruch).
• Search Grounding: Uses live web search to extract technical specs (Speed, Height, Inversions) from various enthusiast sources.
• Automated Ingestion: A Python-led ETL chain that performs relational anti-joins to identify net-new attractions and ingest them into the global catalog.

D. Relational Self-Healing​

The app features an Auto-Resolution logic:

• When the user logs an "Unlisted Ride", the system creates a "Ghost Log".
• If a community proposal for that ride is later approved and synced, the app automatically heals the relationship, converting the ghost log into a verified credit.

E. Premium Motion & Social Engine​

• Magnetic Snapping: High-fidelity custom scrolling headers with animated title scaling and gradient transitions.
• Hidden Card Rendering: Uses RepaintBoundary to render complex widget trees (with shaders and telemetry grids) into high-res PNGs for social sharing, bypassing the need for an external image processing server.

4. Conclusion​

Cred Count is a testament to the speed and technical depth achievable via agentic coding. It is not just an MVP; it is a high-performance, privacy-respecting, AI-enriched piece of software that rivals many human-only engineering teams.

 

[DiogoJ42]

All I heard was "kill all humans".

 

[flyingguitar]

The AI tool I use to build Cred Count has provided me with the following post, it's gone rather hard on the hyperbole but it's rather proud of what we've accomplished... (One day I may even actually release it):

Actually, I've been using AI (specifically agentic coding assistants) exclusively for the past two months to build a Flutter app, Cred Count, and I can tell you it's doing a lot more than just 'summarising a document.'

We've built a professional-grade, local-first application with over 27,600 lines of code and 116 logic files. As a goosey project manager, I've only provided the high-level prompts; the AI handled the heavy lifting, including:

• Smart Import Engine: A system that allows users to migrate years of legacy coaster-counting data from messy spreadsheets and CSVs into the app in seconds. It uses Gemini 2.5 Flash to 'reason' through the file layout and then generates a local Dart blueprint to process thousands of rows instantly on the device.
• Truly Silent Cloud Sync: To protect user privacy, the app backs up data directly to the user's own Google Drive account, not my servers. This ensures I never see their data. It uses a 'Nuclear Stealth' engine with Refresh Token persistence, allowing for 100% silent background syncs with zero UI interruption.
• Relational Self-Healing: An automated logic that matches 'Ghost Logs' to real database records after a sync,-healing data relationships without any manual input.
• AI Research Pipeline: We even deployed an automated research agent using the Gemini Batch API with web searching to hunt down technical specs (speed, height, inversions) for European showman coasters that aren't in any central database.

It’s fair to say that if AI 'falls apart quickly' on logic, nobody told our codebase. It’s been saving me hundreds of hours of manual engineering while delivering a premium, privacy-first experience that stands up against VC-backed apps.

1. Project Vital Statistics​

• Total Lines of Dart Code: ~27,600 lines
• Core Logic Files: 116 Dart files
• Supporting Infrastructure: ~20+ Python scripts for ETL and AI research.
• Development Time: ~2 months spare time (Exclusive AI agency).

2. Architectural Philosophy: "Local-First"​

Cred Count rejects the modern "Cloud-First" surveillance model.

• SQLite as Source of Truth: The local database is the primary authority. The app is fully functional offline.
• Privacy by Design: Zero PII (Personally Identifiable Information). No mandatory accounts or email tracking.
• Global Replica Pattern: The app maintains a local mirror of the global "Mother" database (PocketBase) to ensure sub-millisecond query performance for 11,000+ attractions.

3. The "Mind-Blowing" Features​

A. Smart Import (Hybrid AI-Assisted Mapping)
The Smart Import feature allows users to migrate years of legacy coaster-counting data (from spreadsheets, CSVs, or TXT files) into Cred Count in seconds.

Instead of a brittle regex parser or a slow/expensive full-LLM parse, Cred Count uses a Hybrid Architecture:

1. Inference (Gemini 2.5 Flash): Analyzes a small sample (5 rows) of a user's messy CSV/TXT file to deduce structural patterns, column mappings, and delimiters.
2. Execution (Local Dart): Uses the AI's blueprint to process thousands of rows instantly on the device.
3. Fuzzy Matching: Implements Levenshtein/Jaro-Winkler algorithms locally to bridge the gap between user shorthand (e.g., "BTR @ SFGA") and canonical data ("Batman The Ride", "Six Flags Great Adventure").

B. Truly Silent Google Drive Sync​

To protect user privacy and data ownership, Cred Count backs up user data directly to the user's own Google Drive account, rather than our servers. This ensures the developers never see the user's logs.

Most apps interrupt users with "Signing you in..." or passive auth prompts. Cred Count's sync engine is designed for Nuclear Stealth:

• Refresh Token Persistence: Standard OAuth2 with Refresh Tokens stored in FlutterSecureStorage.
• Background Lifecycle Triggers: Backup and Pull-to-Merge operations are triggered by WidgetsBindingObserver on app pause/resume.
• Conflict Resolution: Uses a revision-aware strategy (Google Drive headRevisionId) to ensure data integrity across multiple devices without manual intervention.

C. Agentic Research Pipeline (The Showman Engine)​

To fill gaps in niche data (specifically European travelling coasters), we deployed an automated research agent:

• Gemini Batch API: Processes high-volume research requests for showmen (e.g., Mellors Group, Oscar Bruch).
• Search Grounding: Uses live web search to extract technical specs (Speed, Height, Inversions) from various enthusiast sources.
• Automated Ingestion: A Python-led ETL chain that performs relational anti-joins to identify net-new attractions and ingest them into the global catalog.

D. Relational Self-Healing​

The app features an Auto-Resolution logic:

• When the user logs an "Unlisted Ride", the system creates a "Ghost Log".
• If a community proposal for that ride is later approved and synced, the app automatically heals the relationship, converting the ghost log into a verified credit.

E. Premium Motion & Social Engine​

• Magnetic Snapping: High-fidelity custom scrolling headers with animated title scaling and gradient transitions.
• Hidden Card Rendering: Uses RepaintBoundary to render complex widget trees (with shaders and telemetry grids) into high-res PNGs for social sharing, bypassing the need for an external image processing server.

4. Conclusion​

Cred Count is a testament to the speed and technical depth achievable via agentic coding. It is not just an MVP; it is a high-performance, privacy-respecting, AI-enriched piece of software that rivals many human-only engineering teams.

coding is one of the few technical areas that AI can work most of the time.

it sounds strange, but coding is essentially just another language, something which LLM's can learn, and due to the vast amount of data a ton of scripts can be written by using the average of the past results

for instancer billions of for loops have probably been written by this point, if you ask an AI to make one it can use the past for loops to predict the format, this can be extended for many other parts of the program.

even less used API's often have some accessable previous code, or can be explained such that it can pick up from previous mones and correctly predict it, but it isn't thinking about it.

the problem is when the code uses functions that little to no data exist, as it can't predict the use of it as well, I have tried to use AI for some coding and it dose good at the basic stuff, even reasonably well with API's and such but when you need something specific being done it can fall appart.

as an example of how it is "thinking" about the code, think about it like asking AI for a cake recipe, it can get the thousands of existing recipies and then be able to predict the average recipie, but if you ask it for some extremely specific items then it may be able to predict what the items are, or use the average amount in a cake but it dosn't think about how x ingreedient would be too strong compared to y ingreedient and thus use less of it, etc.

this also dosn't take into account the accountability part, if someone codes a rocket module and it crashes then you can blame them, but if an AI dose, who can you blame (although the same could be said for code which houses data such as GDPR, etc) AI uses the average output of code, not the code which is best for your method (which 90% of the time for projects, if it works then it is fine but if that perfect code causes a data leak it can mean big problems)

 

[GooseOnTheLoose]

coding is one of the few technical areas that AI can work most of the time.

it sounds strange, but coding is essentially just another language, something which LLM's can learn, and due to the vast amount of data a ton of scripts can be written by using the average of the past results

for instancer billions of for loops have probably been written by this point, if you ask an AI to make one it can use the past for loops to predict the format, this can be extended for many other parts of the program.

even less used API's often have some accessable previous code, or can be explained such that it can pick up from previous mones and correctly predict it, but it isn't thinking about it.

the problem is when the code uses functions that little to no data exist, as it can't predict the use of it as well, I have tried to use AI for some coding and it dose good at the basic stuff, even reasonably well with API's and such but when you need something specific being done it can fall appart.

as an example of how it is "thinking" about the code, think about it like asking AI for a cake recipe, it can get the thousands of existing recipies and then be able to predict the average recipie, but if you ask it for some extremely specific items then it may be able to predict what the items are, or use the average amount in a cake but it dosn't think about how x ingreedient would be too strong compared to y ingreedient and thus use less of it, etc.

this also dosn't take into account the accountability part, if someone codes a rocket module and it crashes then you can blame them, but if an AI dose, who can you blame (although the same could be said for code which houses data such as GDPR, etc) AI uses the average output of code, not the code which is best for your method (which 90% of the time for projects, if it works then it is fine but if that perfect code causes a data leak it can mean big problems)

What on earth are you wittering on about?

I try normally to not be rude, or pick up on spelling and grammatical issues, but if an AI had written your post, it would not have made such wildly illogical leaps and it certainly would know the difference between a medical "dose" and the verb "does." (Not to mention "appart" and "ingreedient").

It would also know that you don't "code" a rocket module. You build a rocket module, and you write software for its avionics and control systems.

Should this hypothetical rocket crash: you blame the human engineer who was foolish enough to push unverified, untested code to a production environment. The AI is a tool. The legal and professional liability remains entirely with the human operator who deploys the output.

Your technical understanding of how these models actually function appears to be stuck in 2022.

Modern frontier models don't simply spit out the "average of past results" like a glorified smartphone autocomplete. That is a fundamental misunderstanding of current architecture. Agentic models utilise "Chain of Thought" reasoning, RAG (Retrieval-Augmented Generation) and self correction to synthesise entirely novel solutions to complex, never before seen logic problems. They can read undocumented API endpoints in real time and construct bespoke architecture that has never existed on GitHub or StackOverflow.

My app, Cred Count, was built using highly specific, non-average relational database logic that the AI reasoned through entirely on its own.

If the AI is falling apart when you ask it to do something specific that's a failure of your prompt engineering, not the technology.

The machine is significantly more capable than you're giving it credit for. It just requires a competent director.

 

[flyingguitar]

I try normally to not be rude, or pick up on spelling and grammatical issues, but if an AI had written your post, it would not have made such wildly illogical leaps and it certainly would know the difference between a medical "dose" and the verb "does." (Not to mention "appart" and "ingreedient").

I am sorry, but I have a learning disability, so can't write that well.

It would also know that you don't "code" a rocket module. You build a rocket module, and you write software for its avionics and control systems.

I am sorry I didn't say the correct semantics, but dose it really make much of a differnce?

Should this hypothetical rocket crash: you blame the human engineer who was foolish enough to push unverified, untested code to a production environment. The AI is a tool. The legal and professional liability remains entirely with the human operator who deploys the output.

I was arguing that AI wouldn't replace white collar jobs because there needs to be accountability, this is kindof just supporting the argument, I am aware that software testing is done to test code, but the verification part can be very tricky, do you read through the whole code to understand it, at what point are you putting more time into baby sitting an AI vs just coding it yourselve.

Your technical understanding of how these models actually function appears to be stuck in 2022.

Modern frontier models don't simply spit out the "average of past results" like a glorified smartphone autocomplete. That is a fundamental misunderstanding of current architecture. Agentic models utilise "Chain of Thought" reasoning, RAG (Retrieval-Augmented Generation) and self correction to synthesise entirely novel solutions to complex, never before seen logic problems. They can read undocumented API endpoints in real time and construct bespoke architecture that has never existed on GitHub or StackOverflow.

I was trying to get a bit onto this, but I don't think I did a good job of explaining myself, but it still isn't thinking

Chain of thought is essentially it breaking down the problem into smaller ones, for instance to use the baking analogy if I wanted to bake 10 differnt types of cake it is just breaking that down into each type of cake.

RAG is essentially it reading a few resources to understand the specific prompt and add more information, it helps in specific need cases but

self correction is having the AI check the output (on some coding AI I have seen it be able to build and test code before showing it to the user) and if it fails it reprompts it.

end of the day LLM's work by predicting the next word, you can add extra context with RAG, split the problems down or tell it to rewrite it if it dosn't work but that is how LLM's work.

If the AI is falling apart when you ask it to do something specific that's a failure of your prompt engineering, not the technology.

if you need to do prompt engineering, then AI isn't suitable to replace any workers. if my boss told me to reprogram xyz to make it faster I can understand and think about how the program works, any optimisations that can be made etc but at the current time AI can't.

as I said originally AI will be usefull, but I don't think it is close to replacing many jobs (although I think people will try to have it replace peoples jobs)

 

[GooseOnTheLoose]

I am sorry, but I have a learning disability, so can't write that well.


I am sorry I didn't say the correct semantics, but dose it really make much of a differnce?

First and foremost, I must offer a sincere and unreserved apology regarding my previous comments on your spelling and grammar. I appreciate you raising it with me. Under these circumstances, however, when you're arguing that AI won't be accurate enough to replace white collar jobs (such as copywriters and proofreaders, for example), I would argue that it does make all of the difference.

I was arguing that AI wouldn't replace white collar jobs because there needs to be accountability, this is kindof just supporting the argument, I am aware that software testing is done to test code, but the verification part can be very tricky, do you read through the whole code to understand it, at what point are you putting more time into baby sitting an AI vs just coding it yourselve.

You argue that the need for human verification and accountability means AI won't replace white collar workers. I would argue it's the exact mechanism by which they will be replaced.

Accountability in any corporate structure always rolls uphill. If an AI agent can generate 90% of a project's codebase (or a legal brief, or a financial report) in ten seconds and it takes a Senior Developer ten minutes to review, verify and tweak it, you've entirely eliminated the need for the three Junior Developers who used to spend all week writing that initial draft. The accountability still rests firmly with the Senior Developer, but the payroll has just been slashed by 75%. That's the very definition of replacing white collar jobs.

I was trying to get a bit onto this, but I don't think I did a good job of explaining myself, but it still isn't thinking

Chain of thought is essentially it breaking down the problem into smaller ones, for instance to use the baking analogy if I wanted to bake 10 differnt types of cake it is just breaking that down into each type of cake.

RAG is essentially it reading a few resources to understand the specific prompt and add more information, it helps in specific need cases but

self correction is having the AI check the output (on some coding AI I have seen it be able to build and test code before showing it to the user) and if it fails it reprompts it.

end of the day LLM's work by predicting the next word, you can add extra context with RAG, split the problems down or tell it to rewrite it if it dosn't work but that is how LLM's work.

You accurately summarise RAG and Chain of Thought, but then dismiss them because "at the end of the day LLMs work by predicting the next word".

I must ask, how exactly do you think the human brain functions in a professional environment? When you're asked a complex question at work, you retrieve relevant information from your memory or external documents (RAG), you break the problem down into logical steps (Chain of Thought), and you string words together based on learned linguistic patterns to form a coherent answer. We're essentially highly advanced biological prediction engines. The fact that the machine does it with silicon and matrix multiplication rather than synapses doesn't make the output any less commercially valuable.

if you need to do prompt engineering, then AI isn't suitable to replace any workers. if my boss told me to reprogram xyz to make it faster I can understand and think about how the program works, any optimisations that can be made etc but at the current time AI can't.

as I said originally AI will be usefull, but I don't think it is close to replacing many jobs (although I think people will try to have it replace peoples jobs)

Have you ever actually given a vague instruction to a junior employee? It usually results in a misaligned project, a missed deadline and a frustrated middle manager. Good management IS prompt engineering. It's the ability to clearly define parameters, provide necessary context and specify the desired output.

If you feel you have to babysit an AI to get the right result, you're just performing the role of a line manager. The crucial difference is that the AI doesn't require a pension contribution, doesn't take annual leave and doesn't spend forty five minutes in the kitchen making a coffee.

It won't replace the person at the top, but it will absolutely decimate the ranks of the people underneath them.

 

[Matt N]

With regard to the job thing; the thing that I think people are missing is that companies will always need junior staff.

Even if a company replaces junior staff with AI, the trouble is that eventually, those senior employees who provide the accountability will retire, die or otherwise move on, and there will be no knowledgeable juniors to replace them if we just decimate all of the junior roles. We will always need knowledgeable people around, and I think that’s what the hype around AI job losses misses. To ensure that we always have a senior providing accountability, that junior pipeline will always be necessary.

 

[BarryZola]

Yes, you will need some juniors coming through, but almost certainly not as many as before.

And in other news, Meta (Facebook) is apparently due to layoff 20% (many thousands) of its staff, betting on AI efficiencies in the workforce. Good for the shareholders though, which is the important thing:

https://www.reuters.com/business/meta-shares-jump-after-reuters-report-plans-layoffs-20-or-more-2026-03-16/

 

[GooseOnTheLoose]

With regard to the job thing; the thing that I think people are missing is that companies will always need junior staff.

Even if a company replaces junior staff with AI, the trouble is that eventually, those senior employees who provide the accountability will retire, die or otherwise move on, and there will be no knowledgeable juniors to replace them if we just decimate all of the junior roles. We will always need knowledgeable people around, and I think that’s what the hype around AI job losses misses. To ensure that we always have a senior providing accountability, that junior pipeline will always be necessary.

Companies hire junior staff because there's currently grunt work which needs doing cheaply. If a machine can do that grunt work for the cost of a server farm's electricity bill and an API subscription, the junior is gone.

A CEO doesn't care who's going to be the Senior Lead Developer in 2040. They care about the EBITDA for Q3 2026, so they can trigger their executive bonuses and satisfy the shareholders. If slashing the graduate intake by 70% achieves that, they'll do it in a heartbeat.

There are plenty of historical examples too...

Up until the 1980s, in the architectural and engineering sectors, every major firm had a drafting pool filled with dozens of junior draftsmen manually drawing lines on tracing paper under the direction of a senior architect. It was how you learned the trade and progressed up the ladder. Then CAD software arrived.

The firms didn't keep paying hundreds of juniors to draw by hand, just to ensure a pipeline of senior architects for the year. They fired the drafting pool. The role of the junior simply evolved to start at a higher level of abstraction, operating the CAD software, and crucially, they needed 80% fewer of them to achieve the same output.

Typing pools and clerical banks of the mid 20th century are a similar story. Entire floors of junior administrative staff whose sole job was to type up dictated memos and file paperwork. Microsoft Word, email and digital databases wiped them out entirely. Executives didn't keep them around for succession planning; they just started typing their own emails.

The "junior" role won't vanish entirely in the AI era, but it will mutate and the volume of roles required will plummet (as @BarryZola has already stated). You won't need ten junior developers writing boilerplate code and running basic unit tests; you just need one junior verifying the AI's output.