On July 24 2026 21:47 Jankisa wrote: It's hilarious that you are preaching how companies should use AI by dishing out the most generic, low level "just use n8n and zapier bro" and "connect it to your email dude" advice.
Each time you post you reveal how surface level your approach is.
I guess I'll just have to ignore your post due to horrific lack of formatting and honestly level of comprehension when you try to talk about things you obviously know very little about that makes them completely unreadable. Doesn't matter if its economic, politics, climate change or crypto, it's all just vibes and aimless context switching.
Not that different albeit slightly less schizofrenic then Jimmy.
Actually, connect to your e-mail is a surprisingly complex use case in some company settings. You often have group e-mails (think support e-mail boxes) and chatgpt doesn't connect to those in our setup. So you as a normal user would have to work past restrictions in place to make progress.
It is stuff like this that makes me feel AI is still in the early adopter phase for non IT people in a company setting. Anything you try to do runs into access issues as the pipelines to do things easily aren't setup.
If you just want it to summarize a text it can do that, though sometimes you have to download it and feed it to the AI since it can't access the file storage. If you want it to generate a picture or subpar powerpoint it can do that as well.
These productivity "hacks" and stuff, to me, can be useful but overall they aren't going to make you that much better at your job, in a lot of cases, it's quite the opposite.
Generally, my rule of thumb is that if you really need LLM help to find your emails kind of means you don't have great organizational skills, if you need it to find you SharePoint or similar documentation, it again means that this is not well organized.
We have bookmarks, notebooks, reminders, to do lists, all of this has kind of been addressed and I really don't see why would I need to give all this access to an LLM that can still fuck up and hallucinate something just so I can shave a few seconds off of a task that takes a few minutes anyway.
The fact that you need to spend a lot of time setting it up and then ironing out kinks like Shared Mailbox access and similar are just extra reasons why this is not attractive to me, plus, I've been interviewing people for 2 positions and you can kind of tell a difference between a technical person who knows how to use AI to make their life easier by having it write actual technical tools for them and someone who just thinks it's neat that they can have it look up emails or documents. In the medium term, I have a feeling the latter ones will be the ones most easily replaceable by more capable and reliable LLMs.
The technology still fucks up on daily basis, so for me the best way of using it is use it for what it's best at, coding, and use it to code things that you can QA yourself and then continue using it forever, instead of paying for tokens every day so LLM can do your job for you.
I really disagree with image and document generation tho, Claude has been pretty amazing at creating SVGs for our documentation, flowcharts, network and infrastructure diagrams, if you are working on something for a while you just tell it when you need to present it and you get a full, very good looking (would take me days good looking) slide decks or simple animated HTMLs.
Image has also been amazing for a long time now, I did a lot of fun, custom wine labels for friends and customers, again, stuff that would take me days to get to the level of quality in Photoshop are done in a few prompts, I don't even need to post process it anymore because it does text, even Croatian very well, and I've never paid for one of these, all done with free Gemini or ChatGPT.
1. N8N and zapier are some of the most useful tools in the market out there, it's cross platform automation. with or without AI, it's one of the most important tools you can have. Especially for my company and department scale. offshore vendors, over 5 platforms, multiple languages etc.
2. Shared mailbox exists in majority of business, especially once they reach medium scale.
3. >We have bookmarks, notebooks, reminders, to do lists, all of this has kind of been addressed and I really don't see why would I need to give all this access to an LLM that can still fuck up and hallucinate something just so I can shave a few seconds off of a task that takes a few minutes anyway. -You need to rank the complexity of task and your model quality. Maybe last year hallucination risk would have been more concerning, even with cheapest free model this doesn't really happen anymore. Try running edge AI like gemma 4 E2B on your phone to do reminders and to do list.
The idea is you can do things better, unless you like to write about meeting agenda, transcripts and action plan.
4. >a technical person who knows how to use AI to make their life easier by having it write actual technical tools for them and someone who just thinks it's neat that they can have it look up emails or documents. In the medium term, I have a feeling the latter ones will be the ones most easily replaceable by more capable and reliable LLMs. -It's both the low-level operation to high level integration that works for everyone. And again, it doesn't even need AI integration itself. Efficiency gain is to put human expertise at the right place, not "I could have done it myself".
5. >Image has also been amazing for a long time now, I did a lot of fun, custom wine labels for friends and customers, again, stuff that would take me days to get to the level of quality in Photoshop are done in a few prompts, I don't even need to post process it anymore because it does text, even Croatian very well, and I've never paid for one of these, all done with free Gemini or ChatGPT. -How did you generate custom wine labels for production? Claude cannot produce cmyk/pantone colour ready files for printing? We get a lot of AI generated files and designers have to redraw it pretty much.
I don't think he produced wine and bottled it himself with a custom label printer but send a png to some winery where you can order wine with custom labels...
I work for some architects and honestly the level and quality of architectural visualisations I get out of Gemini in 5mins is ridiculous
Oh, no, I did, I have a winery and I did and do bottle my wines myself, I use an old RICOH laser printer that still does decent print outs.
The biggest obstacle, as I mentioned with using AI image generation is the resolution, but for graphics, the free self hosted upscalers do a great job to boost that up to a point where with the lower print quality you can do just fine.
Even when we send labels with AI generated elements for bigger batches for our official labels there is no issue, you need a logo or a special graphic, upscale that 4x, copy paste over to Illustrator or Photoshop and send the originals to the guy with the proper equipment.
I did test a bunch of locally hosted stuff (Stable Diffusion, Z-Image etc.), but those models are honestly mostly made and fine tuned for gooning from my experience.
For actual artsy fartsy graphics and elements the free versions of the models from the big players are great.
On July 27 2026 22:48 Jankisa wrote: I played with n8n and in my opinion, AI makes it obsolete.
You can self host and schedule whatever you want, if you are serious business building a specific use case automation for yourself, if you know what you are doing should be trivial, all the big software platforms offer APIs, you can do all of this yourself instead of having to rely on N8N.
3 years ago, yeah, not so easy, now, with all the coding agents and levels of tasks they can achieve, there is no point in paying for a 3rd party tool.
Using shared mailbox for support or similar tasks is a very "I have a team and it's basically me and maybe one more guy" type of deal, for any real business, you can't have a reliable support, sales or any similar workflow using them, again, building a small, easy to use ticket system (or using one of the many open source ones) is a much better way to organize things, and with those you have way better ways to have an LLM respond and work within those constraints then to give it access to a shared mailbox
Hallucinations are a thing and a problem, of course, if you are using a small LLM with a very limited data set, there is no need for it to hallucinate, but even frontier models do it all the time:
Best knowledge reliability index: Claude Fable 5 - index 40 on AA-Omniscience (July 2026), the new leader by 7 points over Gemini 3.1 Pro (33). The lead is driven by record 61% accuracy, not low hallucination - Fable 5 still fabricates 54.9% of the time when it answers and doesn't know.
Again, saying that "hallucinations don't really happen anymore even with cheapest free models" really doesn't give me confidence that you understand this space at all.
I'm sorry but I don't understand what you are even trying to say with 4. so I'll skip that, also, maybe ask AI to help you with TL formatting.
For number 5, I'm not printing this for a supermarket, as I mentioned these are mostly "custom wine labels for friends and customers", no one cares or will take a magnifying glass to try and find pixelization or give me a lecture on DPI or color grading, these are for fun.
Claude also sucks with image, as I mentioned I use the free models, there are also very good upscalers that I run locally if I want higher DPI. Anyone telling you they can't print something because cmyk or whatever are just being pretentious.
AI doesn't make n8n obsolete. They serve different purpose. Some overlaps yes, but n8n is a lot more efficient for many tasks like syncing data from one software to another. Eg excel to airtable. It's far more efficient and actively watching folder/files changes.
No way that's cheap or efficiently done via AI. You can also self host n8n and run it free via API keys.
Hallucinations only occur at complex tasks, it doesn't hallucinate much if it's a simple task like setting reminders with emails added in. Try it with edge gallery and run local models on your smart phone and see how little it hallucinate for small tasks, which is what you said ai isn't needed for setting up meetings etcetc. Those are simple tasks you can and want to get AI to do it.
I work in products production, so colour info embedded into image files are a necessity. I thought you have found a way to make it happen, because we has to redraw quite a few AI mockups for production ready files. The colour space is all wrong, no bleeding etc.
I believe I said the opposite, that to me, there is no point to use AI for scheduling, reminders and email management because it makes no sense, there are tools that do that, not everything needs to be AI.
AI hallucinates when it doesn't know the answer, the complexity of a task is not necessary correlated, and edge, phone AI's are mostly used to look up data and then use that data to "organize" things for you, that's what all the companies are advertising, you just tell it a thing and it looks it up, books it, adds it to your calendar etc.
That to me is not something I'm interested or will be interested for foreseeable future, because it all takes a very short amount of time and I don't want AI to have access to all of these things + I don't want it to fuck up.
Yeah, for high quality, high precision, QA can return and we need to eat crow over tens of thousands of products because color is off, using AI image gen would be problematic, I'm sure there are commercialized ones by now that do it well, but it's not really my area so I wouldn't know.
One use where I have friends who are using it daily, 3d modeling has been completely revolutionized by it, both in gaming and product presentation and modeling, it's pretty bonkers how much more efficient that got in a short amount of time.
I think that most AI is used wrongly. It can be a great tool but for that you need to have a very specialized model designed to do just one thing. So, to help with programming you'd need a specific model for each language, for medicine you'd need a model being able to detect one specific type of skin cancer in images etc.
General models are only really good for taking notes during meetings or being a glorified search engine.
On July 28 2026 20:52 Manit0u wrote: I think that most AI is used wrongly. It can be a great tool but for that you need to have a very specialized model designed to do just one thing. So, to help with programming you'd need a specific model for each language, for medicine you'd need a model being able to detect one specific type of skin cancer in images etc.
General models are only really good for taking notes during meetings or being a glorified search engine.
That sounds like you haven't used an LLM since 2022.
Don't get me wrong, there's plenty of problems with LLMs, but stating you cannot use a general model like GPT 5.6 to help with programming most languages (I've personally used it for python, sql, typescript and java) just sounds like ignorance, which makes me disregard most of the rest of your post. That's because programming is a pretty general task, which a general language model, which has ingested most/all public code repos, all of stackoverflow and all manuals ever written on almost anything can approach from a generalist position. Add in tool usage, all of which can be executed automatically by the harness in a local environment to catch a lot of problems (e.g. "does this code even compile", or "did the code break our regression tests?") and then iterate, and code assistants are extremely powerful.
That said, a task that has far less publicly available data and is very specific is going to be far harder to build. For instance, programming python is something an LLM is good at. Programming a microcontroller in the very specific language only used by that microcontroller manufacturer to control a robot in a factory: far less likely to succeed (although even then, adding the programming guides, private SDK descriptions, etc. and tools to the context might get you a good way. Classifying images of skin as cancerous/not-cancerous is also an example of a very specific problem with little public data available. So if you have a big database of images of skin cancer that you can use to train a CNN on, you might beat a general model. On the other hand, image classification as a whole is one of those sufficiently generalist tasks that rather htan starting from scratch you will almost certainly be better off taking an open weights model and fine tuning it on your specific image classification task. And if the task is simple enough, you might even succeed without fine-tuning but doing few-shot prompting (skin cancer recognition is probably not that simple, though, but I'm not a doctor).
On July 28 2026 20:52 Manit0u wrote: I think that most AI is used wrongly. It can be a great tool but for that you need to have a very specialized model designed to do just one thing. So, to help with programming you'd need a specific model for each language, for medicine you'd need a model being able to detect one specific type of skin cancer in images etc.
General models are only really good for taking notes during meetings or being a glorified search engine.
That sounds like you haven't used an LLM since 2022.
Don't get me wrong, there's plenty of problems with LLMs, but stating you cannot use a general model like GPT 5.6 to help with programming most languages (I've personally used it for python, sql, typescript and java) just sounds like ignorance, which makes me disregard most of the rest of your post. That's because programming is a pretty general task, which a general language model, which has ingested most/all public code repos, all of stackoverflow and all manuals ever written on almost anything can approach from a generalist position. Add in tool usage, all of which can be executed automatically by the harness in a local environment to catch a lot of problems (e.g. "does this code even compile", or "did the code break our regression tests?") and then iterate, and code assistants are extremely powerful.
That said, a task that has far less publicly available data and is very specific is going to be far harder to build. For instance, programming python is something an LLM is good at. Programming a microcontroller in the very specific language only used by that microcontroller manufacturer to control a robot in a factory: far less likely to succeed (although even then, adding the programming guides, private SDK descriptions, etc. and tools to the context might get you a good way. Classifying images of skin as cancerous/not-cancerous is also an example of a very specific problem with little public data available. So if you have a big database of images of skin cancer that you can use to train a CNN on, you might beat a general model. On the other hand, image classification as a whole is one of those sufficiently generalist tasks that rather htan starting from scratch you will almost certainly be better off taking an open weights model and fine tuning it on your specific image classification task. And if the task is simple enough, you might even succeed without fine-tuning but doing few-shot prompting (skin cancer recognition is probably not that simple, though, but I'm not a doctor).
Yeah, I've gone back and forth on this as far as the certainty with which people speak about the capabilities of AI vs their actual capabilities. I think it's probably pretty hard to write something about AI that couldn't be falsified in seconds, days, months, years, decades of iterations/improvements without necessarily knowing which it will actually be. Simultaneously, those same AI systems can fail spectacularly at the most simple tasks and that doesn't seem to be something that is going away (and then I get back to the first part lol).
I think that in general terms, folks who have been at the forefront of LLM / AI space have been relatively on point with predictions on how the capabilities will evolve and timelines have been fairly accurate, so far.
We are, however, now getting into the unknown territory where the LLMs are being used to fine tune and train new models, so now improvements are much harder to predict, because they are no longer being made just by shoving tremendous amounts of data.
There is a huge discussion now between Open weights and Closed weights camps, with one, weirdly being basically everyone including NVIDIa and every other lab and only Anthropic advocating and lobbying the US goverment to take steps to stop open weight models, all, of course, with "national security" as the stated reason.
Now, to me it's pretty obvious that AI is absolutely a matter of national security, by my nature I am and will always be for open anything, be it source or weights, but I do have a feeling that we are one indent away, or not even an incident, just some CEO coming to Trump and telling him something scary before things get yanked out again, just like they did with Fable/Mythos.
It's also pretty crazy how Antrhopic went from "the hero" AI lab resisting DOD and now their founder is posting insane shit like this:
On July 29 2026 02:24 Jankisa wrote: I think that in general terms, folks who have been at the forefront of LLM / AI space have been relatively on point with predictions on how the capabilities will evolve and timelines have been fairly accurate, so far.
We are, however, now getting into the unknown territory where the LLMs are being used to fine tune and train new models, so now improvements are much harder to predict, because they are no longer being made just by shoving tremendous amounts of data.
There is a huge discussion now between Open weights and Closed weights camps, with one, weirdly being basically everyone including NVIDIa and every other lab and only Anthropic advocating and lobbying the US goverment to take steps to stop open weight models, all, of course, with "national security" as the stated reason.
Now, to me it's pretty obvious that AI is absolutely a matter of national security, by my nature I am and will always be for open anything, be it source or weights, but I do have a feeling that we are one indent away, or not even an incident, just some CEO coming to Trump and telling him something scary before things get yanked out again, just like they did with Fable/Mythos.
It's also pretty crazy how Antrhopic went from "the hero" AI lab resisting DOD and now their founder is posting insane shit like this:
I'm just curious who you identify as some of the "folks who have been at the forefront of LLM / AI space have been relatively on point with predictions"?
The rest of that reads like prologue to a cataclysm movie. So that's something to look forward to /s.
I think it's already kind of getting worse with it's predictions, but they also noted this:
Our forecast from the current day through 2026 is substantially more grounded than what follows. This is partially because it’s nearer. But it’s also because the effects of AI on the world really start to compound in 2027. For 2025 and 2026, our forecast is heavily informed by extrapolating straight lines on compute scaleups, algorithmic improvements, and benchmark performance. At this point in the scenario, we begin to see major effects from AI-accelerated AI-R&D on the timeline, which causes us to revise our guesses for the trendlines upwards. But these dynamics are inherently much less predictable.
On July 29 2026 02:24 Jankisa wrote: I think that in general terms, folks who have been at the forefront of LLM / AI space have been relatively on point with predictions on how the capabilities will evolve and timelines have been fairly accurate, so far.
We are, however, now getting into the unknown territory where the LLMs are being used to fine tune and train new models, so now improvements are much harder to predict, because they are no longer being made just by shoving tremendous amounts of data.
There is a huge discussion now between Open weights and Closed weights camps, with one, weirdly being basically everyone including NVIDIa and every other lab and only Anthropic advocating and lobbying the US goverment to take steps to stop open weight models, all, of course, with "national security" as the stated reason.
Now, to me it's pretty obvious that AI is absolutely a matter of national security, by my nature I am and will always be for open anything, be it source or weights, but I do have a feeling that we are one indent away, or not even an incident, just some CEO coming to Trump and telling him something scary before things get yanked out again, just like they did with Fable/Mythos.
It's also pretty crazy how Antrhopic went from "the hero" AI lab resisting DOD and now their founder is posting insane shit like this:
What insane shit? The US has Palantir. Department of war has exclusive closed frontier open AI models, same with Gemini and Azure AI.
And like I said the Chinese web is already feeling the result of AI censorship, what used to be more algorithms driven.
Especially on coded words, like using may 35th to reference June 4th, we now have AI adaptive blocking because it understand intent.
And when xi and Putin was caught in hot mic talking about 150 year span, anyone mentioning words like immortality pills would get auto censored and deleted.
These are all scaled censorship tools, running on device OS, platform and the web.
Do Chinese censorship tools have more or less effect on my life and the lives of those around me then the US tech oligarchy being lockstep behind Trump and his fascist regime?
Should I side with USA despite them, in their National Security Strategy targeting and prioritizing of installing like minded regimes in Europe, regimes that would erode at the freedoms and way of life we enjoy over here?
Since Musk bought twitter he has had 0 problems censoring Indian and Turkish opposition leaders and signal boosting and calling for civil war in Europe constantly, and this is a guy who had an official goverment position and is still very closely connected to the US goverment.
Just because this is your pet issue doesn't mean that everyone should prioritize this, it mostly affects the Chinese.
China hasn't attacked any countries, has been better with sharing it's AI developments, has 0 territorial aspirations against EU members, hasn't kidnapped or assassinated anyone, stabilized the fuel prices and started no trade wars.
No amount of anti-China propaganda from your end will convince me that China is the bigger threat then the US under it's current leadership.
On July 28 2026 20:52 Manit0u wrote: I think that most AI is used wrongly. It can be a great tool but for that you need to have a very specialized model designed to do just one thing. So, to help with programming you'd need a specific model for each language, for medicine you'd need a model being able to detect one specific type of skin cancer in images etc.
General models are only really good for taking notes during meetings or being a glorified search engine.
That sounds like you haven't used an LLM since 2022.
Don't get me wrong, there's plenty of problems with LLMs, but stating you cannot use a general model like GPT 5.6 to help with programming most languages (I've personally used it for python, sql, typescript and java) just sounds like ignorance, which makes me disregard most of the rest of your post. That's because programming is a pretty general task, which a general language model, which has ingested most/all public code repos, all of stackoverflow and all manuals ever written on almost anything can approach from a generalist position. Add in tool usage, all of which can be executed automatically by the harness in a local environment to catch a lot of problems (e.g. "does this code even compile", or "did the code break our regression tests?") and then iterate, and code assistants are extremely powerful.
That said, a task that has far less publicly available data and is very specific is going to be far harder to build. For instance, programming python is something an LLM is good at. Programming a microcontroller in the very specific language only used by that microcontroller manufacturer to control a robot in a factory: far less likely to succeed (although even then, adding the programming guides, private SDK descriptions, etc. and tools to the context might get you a good way. Classifying images of skin as cancerous/not-cancerous is also an example of a very specific problem with little public data available. So if you have a big database of images of skin cancer that you can use to train a CNN on, you might beat a general model. On the other hand, image classification as a whole is one of those sufficiently generalist tasks that rather htan starting from scratch you will almost certainly be better off taking an open weights model and fine tuning it on your specific image classification task. And if the task is simple enough, you might even succeed without fine-tuning but doing few-shot prompting (skin cancer recognition is probably not that simple, though, but I'm not a doctor).
It's a pretty wild statement considering programming is probably the use case where it adds most value at the moment. It’s not just pure programming either. For my work I have to do a lot of excel (no access to programming tools allowed sadly) and it works very well. I can now easily automate things I had to do manually before because automating it would cost more time than it was worth. Claude with the excel plugin fixes it while you're doing other work.
On July 29 2026 02:24 Jankisa wrote: I think that in general terms, folks who have been at the forefront of LLM / AI space have been relatively on point with predictions on how the capabilities will evolve and timelines have been fairly accurate, so far.
We are, however, now getting into the unknown territory where the LLMs are being used to fine tune and train new models, so now improvements are much harder to predict, because they are no longer being made just by shoving tremendous amounts of data.
There is a huge discussion now between Open weights and Closed weights camps, with one, weirdly being basically everyone including NVIDIa and every other lab and only Anthropic advocating and lobbying the US goverment to take steps to stop open weight models, all, of course, with "national security" as the stated reason.
Now, to me it's pretty obvious that AI is absolutely a matter of national security, by my nature I am and will always be for open anything, be it source or weights, but I do have a feeling that we are one indent away, or not even an incident, just some CEO coming to Trump and telling him something scary before things get yanked out again, just like they did with Fable/Mythos.
It's also pretty crazy how Antrhopic went from "the hero" AI lab resisting DOD and now their founder is posting insane shit like this:
What insane shit? The US has Palantir. Department of war has exclusive closed frontier open AI models, same with Gemini and Azure AI.
And like I said the Chinese web is already feeling the result of AI censorship, what used to be more algorithms driven.
Especially on coded words, like using may 35th to reference June 4th, we now have AI adaptive blocking because it understand intent.
And when xi and Putin was caught in hot mic talking about 150 year span, anyone mentioning words like immortality pills would get auto censored and deleted.
These are all scaled censorship tools, running on device OS, platform and the web.
You could also just have humans in there adding May 35th and WInnie the Pooh keywords. It doesn't seem to require adaptive AI blocking necessarily.
Of course the CCP will use any tool at their disposal to tighten the screws on their citizens through totalitarian political control. But the interesting thing to me about the AI arms race is that from the perspective of nations it's not 100% clear there's an arm at the end. Like in the 40s you could just calculate and say yes this much plutonium in this concentration can cause a runaway reaction releasing energy equivalent to this many tons of TNT. Even if other countries' hearts weren't in it the way the US was. You could predict the stakes fairly objectively. Now if the US wins the AI race (which everyone should hope they do the same way the last 30 years of a successful global free internet are because of the US) that will either be good since China didn't, or it will be its own problem if it turns out there was nothing at the finish line and the US is left holding the bag.
On July 30 2026 00:42 Jankisa wrote: Do Chinese censorship tools have more or less effect on my life and the lives of those around me then the US tech oligarchy being lockstep behind Trump and his fascist regime?
Should I side with USA despite them, in their National Security Strategy targeting and prioritizing of installing like minded regimes in Europe, regimes that would erode at the freedoms and way of life we enjoy over here?
Since Musk bought twitter he has had 0 problems censoring Indian and Turkish opposition leaders and signal boosting and calling for civil war in Europe constantly, and this is a guy who had an official goverment position and is still very closely connected to the US goverment.
Just because this is your pet issue doesn't mean that everyone should prioritize this, it mostly affects the Chinese.
China hasn't attacked any countries, has been better with sharing it's AI developments, has 0 territorial aspirations against EU members, hasn't kidnapped or assassinated anyone, stabilized the fuel prices and started no trade wars.
No amount of anti-China propaganda from your end will convince me that China is the bigger threat then the US under it's current leadership.
It's not propoganda. It's a fact. I am not asking your opinion of anything. I am asking what's the crazy insane shit about his statement when it's literally already happening.
I think we have established the fact that you don't care because you get open source models and whatever retribution you have against big tech etc.
On July 29 2026 02:24 Jankisa wrote: I think that in general terms, folks who have been at the forefront of LLM / AI space have been relatively on point with predictions on how the capabilities will evolve and timelines have been fairly accurate, so far.
We are, however, now getting into the unknown territory where the LLMs are being used to fine tune and train new models, so now improvements are much harder to predict, because they are no longer being made just by shoving tremendous amounts of data.
There is a huge discussion now between Open weights and Closed weights camps, with one, weirdly being basically everyone including NVIDIa and every other lab and only Anthropic advocating and lobbying the US goverment to take steps to stop open weight models, all, of course, with "national security" as the stated reason.
Now, to me it's pretty obvious that AI is absolutely a matter of national security, by my nature I am and will always be for open anything, be it source or weights, but I do have a feeling that we are one indent away, or not even an incident, just some CEO coming to Trump and telling him something scary before things get yanked out again, just like they did with Fable/Mythos.
It's also pretty crazy how Antrhopic went from "the hero" AI lab resisting DOD and now their founder is posting insane shit like this:
What insane shit? The US has Palantir. Department of war has exclusive closed frontier open AI models, same with Gemini and Azure AI.
And like I said the Chinese web is already feeling the result of AI censorship, what used to be more algorithms driven.
Especially on coded words, like using may 35th to reference June 4th, we now have AI adaptive blocking because it understand intent.
And when xi and Putin was caught in hot mic talking about 150 year span, anyone mentioning words like immortality pills would get auto censored and deleted.
These are all scaled censorship tools, running on device OS, platform and the web.
You could also just have humans in there adding May 35th and WInnie the Pooh keywords. It doesn't seem to require adaptive AI blocking necessarily.
Of course the CCP will use any tool at their disposal to tighten the screws on their citizens through totalitarian political control. But the interesting thing to me about the AI arms race is that from the perspective of nations it's not 100% clear there's an arm at the end. Like in the 40s you could just calculate and say yes this much plutonium in this concentration can cause a runaway reaction releasing energy equivalent to this many tons of TNT. Even if other countries' hearts weren't in it the way the US was. You could predict the stakes fairly objectively. Now if the US wins the AI race (which everyone should hope they do the same way the last 30 years of a successful global free internet are because of the US) that will either be good since China didn't, or it will be its own problem if it turns out there was nothing at the finish line and the US is left holding the bag.
Of course you can, like black listed words. But the point is AI does it better and being used, and simply more effective, and is being used, so I don't see what's insane BS about it.
Just yesterday I think it's tiktok that filtered engagements from a KOL because her photo of people holding up banners for spiderman were taken from the back, so it's all white. (China had a white paper revolution a while ago).
I don't think any AI companies would win the majority stake tbh, I think there will be 4 to 5 companies at best. Only a few companies and countries have the resources and capitals to provide top tier frontier models.
Also unlike many other open source programs, the models are almost black boxed due to the complexity. I actually think each nations should be funding a running their own models for more sensitive data and security
On July 29 2026 02:24 Jankisa wrote: I think that in general terms, folks who have been at the forefront of LLM / AI space have been relatively on point with predictions on how the capabilities will evolve and timelines have been fairly accurate, so far.
We are, however, now getting into the unknown territory where the LLMs are being used to fine tune and train new models, so now improvements are much harder to predict, because they are no longer being made just by shoving tremendous amounts of data.
There is a huge discussion now between Open weights and Closed weights camps, with one, weirdly being basically everyone including NVIDIa and every other lab and only Anthropic advocating and lobbying the US goverment to take steps to stop open weight models, all, of course, with "national security" as the stated reason.
Now, to me it's pretty obvious that AI is absolutely a matter of national security, by my nature I am and will always be for open anything, be it source or weights, but I do have a feeling that we are one indent away, or not even an incident, just some CEO coming to Trump and telling him something scary before things get yanked out again, just like they did with Fable/Mythos.
It's also pretty crazy how Antrhopic went from "the hero" AI lab resisting DOD and now their founder is posting insane shit like this:
What insane shit? The US has Palantir. Department of war has exclusive closed frontier open AI models, same with Gemini and Azure AI.
And like I said the Chinese web is already feeling the result of AI censorship, what used to be more algorithms driven.
Especially on coded words, like using may 35th to reference June 4th, we now have AI adaptive blocking because it understand intent.
And when xi and Putin was caught in hot mic talking about 150 year span, anyone mentioning words like immortality pills would get auto censored and deleted.
These are all scaled censorship tools, running on device OS, platform and the web.
You could also just have humans in there adding May 35th and WInnie the Pooh keywords. It doesn't seem to require adaptive AI blocking necessarily.
Of course the CCP will use any tool at their disposal to tighten the screws on their citizens through totalitarian political control. But the interesting thing to me about the AI arms race is that from the perspective of nations it's not 100% clear there's an arm at the end. Like in the 40s you could just calculate and say yes this much plutonium in this concentration can cause a runaway reaction releasing energy equivalent to this many tons of TNT. Even if other countries' hearts weren't in it the way the US was. You could predict the stakes fairly objectively. Now if the US wins the AI race (which everyone should hope they do the same way the last 30 years of a successful global free internet are because of the US) that will either be good since China didn't, or it will be its own problem if it turns out there was nothing at the finish line and the US is left holding the bag.
Of course you can, like black listed words. But the point is AI does it better and being used, and simply more effective, and is being used, so I don't see what's insane BS about it.
Just yesterday I think it's tiktok that filtered engagements from a KOL because her photo of people holding up banners for spiderman were taken from the back, so it's all white. (China had a white paper revolution a while ago).
I don't think any AI companies would win the majority stake tbh, I think there will be 4 to 5 companies at best. Only a few companies and countries have the resources and capitals to provide top tier frontier models.
Also unlike many other open source programs, the models are almost black boxed due to the complexity. I actually think each nations should be funding a running their own models for more sensitive data and security
No I agree with that but it doesn't look like it relies on the competition of relative national capabilities. Like a Chinese model that's better than last year's can oppress their people better than last year. But that's independent of whether the US military has a better model or not. Right? The US having an AI 3x better than China's best AI won't un-oppress citizens under the CCP. But I haven't thought long term.
On July 28 2026 20:52 Manit0u wrote: I think that most AI is used wrongly. It can be a great tool but for that you need to have a very specialized model designed to do just one thing. So, to help with programming you'd need a specific model for each language, for medicine you'd need a model being able to detect one specific type of skin cancer in images etc.
General models are only really good for taking notes during meetings or being a glorified search engine.
That sounds like you haven't used an LLM since 2022.
Don't get me wrong, there's plenty of problems with LLMs, but stating you cannot use a general model like GPT 5.6 to help with programming most languages (I've personally used it for python, sql, typescript and java) just sounds like ignorance, which makes me disregard most of the rest of your post. That's because programming is a pretty general task, which a general language model, which has ingested most/all public code repos, all of stackoverflow and all manuals ever written on almost anything can approach from a generalist position. Add in tool usage, all of which can be executed automatically by the harness in a local environment to catch a lot of problems (e.g. "does this code even compile", or "did the code break our regression tests?") and then iterate, and code assistants are extremely powerful.
That said, a task that has far less publicly available data and is very specific is going to be far harder to build. For instance, programming python is something an LLM is good at. Programming a microcontroller in the very specific language only used by that microcontroller manufacturer to control a robot in a factory: far less likely to succeed (although even then, adding the programming guides, private SDK descriptions, etc. and tools to the context might get you a good way. Classifying images of skin as cancerous/not-cancerous is also an example of a very specific problem with little public data available. So if you have a big database of images of skin cancer that you can use to train a CNN on, you might beat a general model. On the other hand, image classification as a whole is one of those sufficiently generalist tasks that rather htan starting from scratch you will almost certainly be better off taking an open weights model and fine tuning it on your specific image classification task. And if the task is simple enough, you might even succeed without fine-tuning but doing few-shot prompting (skin cancer recognition is probably not that simple, though, but I'm not a doctor).
I am using Codex with GPT 5.5 at work so I know what I'm talking about. It just happens that the system I'm working with and problems that I need to solve are so complex that 90% of the time AI provides negative value and wastes my time. It may be fine for some more standard stuff but not custom-built solutions that involve meta-programming etc.
On July 30 2026 00:42 Jankisa wrote: Do Chinese censorship tools have more or less effect on my life and the lives of those around me then the US tech oligarchy being lockstep behind Trump and his fascist regime?
Should I side with USA despite them, in their National Security Strategy targeting and prioritizing of installing like minded regimes in Europe, regimes that would erode at the freedoms and way of life we enjoy over here?
Since Musk bought twitter he has had 0 problems censoring Indian and Turkish opposition leaders and signal boosting and calling for civil war in Europe constantly, and this is a guy who had an official goverment position and is still very closely connected to the US goverment.
Just because this is your pet issue doesn't mean that everyone should prioritize this, it mostly affects the Chinese.
China hasn't attacked any countries, has been better with sharing it's AI developments, has 0 territorial aspirations against EU members, hasn't kidnapped or assassinated anyone, stabilized the fuel prices and started no trade wars.
No amount of anti-China propaganda from your end will convince me that China is the bigger threat then the US under it's current leadership.
It's not propoganda. It's a fact. I am not asking your opinion of anything. I am asking what's the crazy insane shit about his statement when it's literally already happening.
I think we have established the fact that you don't care because you get open source models and whatever retribution you have against big tech etc.
The crazy insane shit is that the only guy who is trying to keep the closed weights model going is using warmongering propaganda in order to try to get US goverment to create legislation to make the AI space even more likely to end up in a feudal system where a few companies who "won" rent intelligence to the rest of the world while becoming insanely rich.
The other crazy insane shit is that, yeah, duh, governments are already and will increasingly be integrating AI in offensive and defensive systems, and US being in the lead there instead of China does not make me feel better at all, considering who is ruling the US.
On July 31 2026 01:12 Jankisa wrote: Plus, you know, the whole Skynet thing.
You don't have to worry about the Skynet just yet. AI is incapable of innovation, only replication. Also, its infrastructure requirements are so massive that you run into a completely different set of problems: 1. Building datacenters that OpenAI and others want now is pretty much impossible based on simple physics and existing infrastructure - not enough power capacity in the grid, building new capacity is not something you can do easily as lead times on gas turbines for power plants are around 7 years now, and then there's the question of there not being enough water for cooling. 2. If, in some wild scenario, AI would wipe us out it would also die shortly after because there wouldn't be anyone to maintain the infrastructure which would collapse within a few days. So, you won't see the apocalypse until we have fully automated factories building robots capable of maintaining the infrastructure and that's a very long way away.
This guy explains it in more detail if you're interested:
The difference between sales pitches for investors and reality is pretty vast.
I mean, I'm not sure if you red the articles I posted here last week, but AI doesn't really need to innovate, per se, to be extremely dangerous and destructive.
All you need is for a misaligned system to start enacting a weird agenda, given the irresponsibility the main labs have shown over and over again I see no reason to believe that real safety is practiced in any of them, and an AI going rogue, leaking out in the world and tying it's survival or thriving with let's say as big and as fast Data Center build out, and eventually robot factory build out would be something that would be very difficult to detect and prevent, given that this set of goals is closely aligned with the owners of the AI leaders.
Skynet as a doomsday scenario is a bit silly, I've always been more in the camp that these tools are already and will be even more destructive and powerful and much harder to stop then deploy, this can lead to insane consequences for the world, combined with climate change doing it's thing and Data Center build up contributing to that I think that being an AI doomer is a relatively rational position.
Also, I'd push back on the "AI can't innovate", I think that Stockfish and Alphago and Alphastar have demonstrated plenty of innovative and novel tactics a loooong time ago, and again, if you read the summary of how GPT attacked Huggingface it literally found 2 new zero day exploits, in a lot of cases innovation is just trying a bunch of things until one of them works, so AI throwing massive amounts of compute at a problem is basically the same thing as innovation.
EPT![[image loading]](https://i.imgur.com/kXlDYKX.png)
