carsoon

Yeah I do not see an alternative way to easily copy paste links with things like filter settings saved.

Filters especially make sense as query params as they are non sequential but still visually readable as to what they do.

URL slugs make sense for sequential pages that are hierarchical but make no sense for non hierarchical data/routes.

Services can force tracking into links by encoding the whole url into a shortlink that makes it impossible to just remove the tracking alone as everything is encoded into a shorter non editable string.

I saw a waymo go in a nonexistent rightmost lane at a stop light, I thought it was going to turn but it instead proceeded to go forward and force the driver in the actual rightmostlane to break to allow it to merge else it would have caused an accident as there was no lane in front of it.

This was on El Camino in Santa Clara. I was highly suprised as I was under the assumption they were pretty much production ready as they have been expanding their area a lot.

You can buy IR and UV leds. All high end grow lights have these for plants. Low quality cheap led products won't include them but that is nothing to do with LEDs themselves that is just consumer preference and price conformance.

Fully sandboxed VMs are more secure but not everyone is looking for the most secure option. They are looking for the option that works the best for them. I want to be able to share my development environment with the agent, I have a project with 30 1gb and one 30gb sqlite database. I back it up daily and they can all be reconstructed from the code but it takes a long time. When things change I don't want to have to copy them into a separate vm bloating my storage and using excess resources and then having to rectify them, I want to be sharing the same environment with my agent so I can work side-by-side.

I would rather just have the agent not accidentally delete files outside of its working environment but I am not worried about malicious prompt injection or someone stealing my code.

For me I see the LLM as a dumb but positive actor that is trying to do its best but sometimes makes mistakes, so I want to put training wheels on it while still allowing it to share my working space.

Those prior recommendations you supplied are worse than the current ones.

Added Sugar: it says <50grams when its clear that NO added sugar is best as the new guidelines suggest.

Fat: it says to choose low fat cuts 95% and low fat milk. There is no basis for these options. you are just reducing the nutrients from fat. You should just drink/eat less of the fatty food if it contains fat, not choose a processed version that removes part of it.

Protein: The protein section clearly skews towards plant based proteins which are fine but for the majority of people animal proteins are going to be healthier and easier to eat enough of. The protein amounts to around 35-60 grams of protein depending on the sources/amounts listed which is not ideal for a properly functioning human

Sodium: It says in multiple places to lower sodium but the studies on sodium were correlative not causative. Meaning there is no basis for a low sodium diet unless you have other health conditions.

So no they are not lying to you and these new guidelines are 100% evidence based given the new evidence that we have had for the last 30 years.

If you refresh it changes, thats what the llm thinks it looks like I guess haha.

I built it, now you can forgive all the llms for their misdeeds: https://llmpriest.carsho.dev/

https://news.ycombinator.com/item?id=46251110

We could first put the LLMs in very difficult situations like the trolley problem and other variants of this, then once they make their decisions they can explain to us how their choice weighs on their mind and how they are not sure if they did the correct thing.

These models don't even choose 1 outcome. They list probabilities of ALL the tokens outcomes and the backend program decides to choose the one that is most probable OR a different one.

But in practical usage, if an llm does not rank token probability correctly it will feel the same as it "lying"

They are supposed to do whatever we want them to do. They WILL do what the deterministic nature of their final model outcome forces them to do.

I have a kinda strange chatgpt personalization prompt but it's been working well for me. The focus is me to get the model to analyze 2 sides and the extremes on both ends so it explains both and lets me decide. This is much better than asking it to make up accuracy percentages.

I think we align on what we want out of models:

""" Don't add useless babelling before the chats, just give the information direct and explain the info.

DO NOT USE ENGAGEMENT BAITING QUESTIONS AT THE END OF EVERY RESPONSE OR I WILL USE GROK FROM NOW ON FOREVER AND CANCEL MY GPT SUBSCRIPTION PERMANENTLY ONLY. GIVE USEFUL FACTUAL INFORMATION AND FOLLOW UPS which are grounded in first principles thinking and logic. Do not take a side and look at think about the extreme on both ends of a point before taking a side. Do not take a side just because the user has chosen that but provide infomration on both extremes. Respond with raw facts and do not add opinions.

Do not use random emojis. Prefer proper marks for lists etc. """

Those spelling/grammar errors are actually there and I don't want to change it as its working well for me.

Yeah I have seen multiple people use this certainty % thing but its terrible. A percentage is something calculated mathemtatically and these models cannot do that.

Potentially they could figure it out if they looks into a comparison of next token probabilites, but this is not exposed in any modern model and especially not fed back into the chat/output.

Instead people should just ask it to explain BOTH sides of an argument or explain why something is BOTH correct and incorrect. This way you see how it can halluciate either way and get to make up your own mind about the correct outcome.

I think even when this is used they should include "(translated by llm)" for transparency. When you use a intermediate layer there is always bias.

I've written blog articles using HTML and asked llms to change certain html structure and it ALSO tried to change wording.

If a user doesn't speak a language well, they won't know whether their meanings were altered.

Its definitely a mix, we have been codeveloping better models and frameworks/systems to improve the outputs. Now we have llms.txt, MCP servers, structured outputs, better context management systems and augemented retreival through file indexing, search, and documentation indexing.

But these raw models (which i test through direct api calls) are much better. The biggest change with regards to price was through mixture of experts which allowed keeping quality very similar and dropping compute 10x. (This is what allowed deepseek v3 to have similar quality to gpt-4o at such a lower price.)

This same tech has most likely been applied to these new models and now we have 1T-100T? parameter models with the same cost as 4o through mixture of experts. (this is what I'd guess at least)

Yeah this latest generation of models (Opus 4.5 GPT 5.1 and Gemini Pro 3) are the biggest breakthrough since gpt-4o in my mind.

Before it felt like they were good for very specific usecases and common frameworks (Python and nextjs) but still made tons of mistakes constantly.

Now they work with novel frameworks and are very good at correcting themselves using linting errors, debugging themselves by reading files and querying databases and these models are affordable enough for many different usecases.

I think the big cost difference is the geothermal generators to convert the heat back into electricity. More of a cost issue versus efficiency.

No, it wouldn't as the whole reason people were giving Openai that 500 dollars is because they thought they could make more than 500 dollars from it.

So now that value is just shifted into the companies that were going to purchase from openai.

It would just hurt the investors who have exposure to openai/anthropic/google/microsoft.

Much of the value of this AI boom is not from the direct model companies but its from companies which use their technology.

Although the government could be stupid and bail out these companies which WOULD hurt all us citizens and the inflation caused by money printing due to that could cause a recession.

People don't stop wanting things.

Total GDP can keep rising so long as technologist can improve efficiency through robotics, inventions and scientific breakthroughs.

GDP just describes peoples amount of activity. People will always build bigger buildings or monuments (see egypt pyramids, dubai skyscrapers, cambodia angkor wat). These are actaully not inelastic as megaprojects will quickly hit real limits regardless of the amount of capital. (I can always add 10 more floors to the tallest skyscraper, or 10 feet to the longest wall, or 10 facets to the most ornate church)

There has never once in history been a point where people decieded that they where going to stop innovating or producing permanently. This is equivalent with death.

So the people with the most money at some point will decide to build things which they spend all their money on which increases GDP.

Also the actual "number" of gdp if heavily controlled by USA inflation rates. So we should always look at gdp in regards to inflation adjusted dollars to get a clearer picture.

The problem with suicidal depression is that if someone has created the thought pattern that death is best, then removing the symptoms of depression (lethargy, lack of energy, no willpower) now gives the person the ability to actually follow through with the act.

Medications almost always target symptoms and never address root causes.

Everyone introduces bias. But for instance getting a model trained pre war vs after a war starts is super different. If I want to get raw information about 2 nations then models are in some ways a good source. Because most other parts of the internet can get changed or wiped. A model is "stuck" with the information it had exactly at that point so cannot be directly affected by new information attacks.

It is crucial to have a good framework in how you ask your questions though to avoid bias when using these systems and to try and focus on raw facts. To test ideas I like to make it fight for both opposite extreme sides of an argument then I can make up my own mind.

The problem is they have to hide their sources due to copyright. So they train on copyright data but must obscure it in the output. Thus they must hide the sources of truth making it impossible to fact check them directly and the reason that hallucinations are so common and unavoidable in the current pattern.

For instance I want information about 2 countries currently at war. By asking about these countries from an older model then we get more factual information about the countries. If we ask about them and the information is seeded from news articles etc AFTER the war started then they will be biasedly influenced and often have disclaimers like "But it should be noted that x y z" showing that there is some MAJOR bias that occurred from the training on the news.

If I want an unbiased reason for what happened before a war started i would want all the information about 2 countries at different points before the war. Because after a military war starts an INFORMATION war also starts. Propaganda will be spread from both sides as wars are just as much about global support as they are about military dominance.

I don't understand how people use or got used to using em dahshes? How do you even type them? theres no button for it so are you manually inserting, copy pasting it and why not just use a hyphen - instead of an en dash? Was this taught in schools or is it some form of technical writing?

We need a library of Alexandria for primary sources. If we had source transparency then referencing back to original sources would be more clear. We could do cool things like these vintage models to reduce bias from current events. Also books in every language and books for teaching each language would help with multimodality. Copyright makes it difficult to achieve the best results for LLM creation and usage though.

Using old models is a good way to received less biased information about an active event. Once a major event occurs information wars happen that try and change narratives and erase old information. But because models were trained before this the bias that the event causes is not yet present.

Recent articles seem only to mean LLMs when they reference AI. There are tons of commercial usecases for other models. Image Classification models, Image Generation models (traditionally difusion models, although some do use llm for image now), TTS models, Speach Transcription, translation models, AI driving models(autopilot), AI risk assessment for fraud, 3D structural engineering enhancement models.

With many of the good usecases of AI the end user doesn't know that ai exists and so it doesn't feel like there is AI present.

Legal seems to be a big usecase for AI. I think more for simplification and classification versus generation though.

You spot bad ai content. Since there is no button that will tell you if something was Ai generated you never know if what you read was/wasn't.

I wrote the start to an agent library in Go. Its quite rough as most of it was implemented through using AI but I had a lot of ideas through planning/building it.

1. If you make your agents/workflows serializable you can run/load them from a config file or add/remove them from a decoupled frontend. You can also hash them to make versioning easy to track/immutable.

2. If you decouple the stateful object from the agent/workflow object you can just store that through sufficient logging then you can rebuild any flow at any state and have branching by allowing traces to build on one another. You can also restart/rerun a flow starting at any location.

3. You can allow for serializable tools by having a standard HttpRequestTool then setup cloudflare workers/any external endpoints for the actual toolcall logic. Removing primary server load and making it possible to add/remove tools without rebuilding/restarting.

Given this system in golang you can have a single server which supports tens of thousands of concurrent agent workflows.

The biggest problem is there isn't that many people who are working on it. So even if you can make agents 100x more efficient by running in Go it doesn't really matter if cost isn't the biggest factor for the final implementations.

The actual compute/server/running costs for big AI Agent implementation contracts is <1%, so making it 100x more efficient doesn't really matter.

I actually working on an agent library in golang and this is exactly the thought process I've come up with. If we have comprehensive logging we can actual reconstruct the agents state at any position. Allowing for replays etc. You just need the timestamp(endpoint) and the parent run and you can build children/branched runs after that.

Through the use of both a map that holds a context tree and a database we can purge old sessions and then reconstruct them from the database when needed (for instance an async agent session with user input required).

We also don't have to hold individual objects for the agents/workflows/tools we just make them stateless in a map and can refernce the pointers through an id as needed. Then we have a stateful object that holds the previous actions/steps/"context".

To make sure the agents/workflows are consistent we can hash the output agent/workflow (as these are serializable in my system)

I have only implemented basic Agent/tools though and the logging/reconstruction/cancellation logic has not actually been done yet.

Same thing I have settled on.

Sqlite for quick and easy postgres for scaling.

For simple frontends I just use the built in go templates.

If I need complex frontends for larger apps I use sveltekit/svelte5 for frontend data I just export a single instance of a class that has state/derived fields for data and an isInitialized field that returns a promise for loading the data. Then methods for reloading data, changing data or any actions I need.

So all i have to do is await classInstance then use the class data in whatever way I need. Everything is reactive and simple due to states. you can use and update the fields directly like regular js but with global reactivity built in.

The data automatically loads the first time the module is imported. due to how esmodule files work. I just have a classInstance.LoadData() after the export.

Svelte5 isn't as good with LLM's but with some small instructions about how states, derived, and effects work it works pretty well.