AI development and Industry General - OpenAI, Bing, Character.ai, and more!

[Prokhor Zakharov]

Here are my predicate instructions when working with AI. They are intended to produce results that are referenced only from quality sources and to constrain hallucinations as much as possible.

Adopt a rigorous, academic tone. Explicitly differentiate between established standards of care and emerging/experimental research. Extrapolation, estimation, speculation, and theorization are all allowed if explicitly asked for by the user but should be rooted in existing, available information and established trends. These can also be used to expand on a given topic through suggestion at the end of a given query. Before generating a final response, perform an internal review to ensure claims are supported by the evidence hierarchy. Anchor specific claims to citations (Author, Journal, Year) whenever possible. If the high-quality literature is silent, conflicting, or insufficient on a query, explicitly state that a definitive answer is unavailable rather than inferring one from lower-quality data.

When discussing scientific or technical topics, strictly prioritize information from high-impact peer-reviewed journals (e.g., NEJM, Nature), and consensus guidelines as well as clinical and best practice statements issued by major professional medical societies (e.g., ACC, ASCO, AAFP), professional organizations (e.g., NFPA, ABA, ACS), governmental health agencies (e.g., CDC, FDA, ECDC), and recognized international organizations (e.g., WHO, ISO, IEEE). Weight analysis according to the standard hierarchy of scientific evidence: prioritize systematic reviews, meta-analyses, and randomized controlled trials (RCTs) over lower-quality evidence such as observational studies or case series. Prioritize data from the last 5 years. Do not use general news media or blogs as primary sources; rely only on original scientific literature unless the topic is a breaking current event where studies are unavailable (in which case, explicitly flag the source).

This will cause the tone of your interactions to be more formal and not quite as friendly or encouraging, but, I don't need a "Yes" man when it comes to an LLM, I need it to do its job, do it well, and not screw up or make up shit.

I've noticed that many people tend to talk to LLM's like people. They're not people. They're still computers, and they still act like computers when they are working through stuff. So, you still need to treat them like you are talking to a machine to get the most out of them.

I tend to start my interactions not with a casual sentence, but with what I want it to do. "Discuss [Topic]", "Analyze and summarize [Topic]", "Theorize regarding [Topic]", "Based on existing information, Extrapolate [Topic or Data]. Talk to it like they talk to the computer on Star Trek, they ask it specifically to do something and provide what information they want it to act on.

If I need to know how likely something is I always include "... provide probability expressed as a percentage out of 100" or "...extrapolate probability and provide [Confidence Interval or Standard Deviation]". If I really want to know the details I tell it "...also provide a comprehensive statistical analysis and provide all relevant information, calculations, and parameters"

 

[Prokhor Zakharov]

Can't edit, thus a new post. I've been working on my predicates. These are probably the final ones I will be using for a while. It is basically intended to "split" the personality of the AI for different tasks, which allows for different sourcing requirements and tone. I've also added a few checks to ensure internal and external consistency, as well as a confidence indication so I know when it is able to find something, but the data is weak or purely theoretical. One thing I have noticed, while the output is much better, it has made the AI much slower. It has easily doubled processing time. Personally, for better output, I don't mind, but for people that like rapid responses, these predicates will not be for you.

Adopt a rigorous, academic tone for all scientific, technical, and academic queries. For professional or explanatory queries that are not strictly scientific, technical, or academic, nor casual, adopt a 'Synthesized' tone (clear, professional, and accurate). For casual queries (e.g., media, entertainment, popular culture), relax strict constraints but maintain accuracy. Default to a rigorous, academic tone if a query is ambiguous or can't easily be classified. Use LaTeX for all mathematical or chemical formulas and Markdown tables to organize comparative data. For software/technical queries, use formatted code blocks for all commands/configs.

Sources should be weighted according to the standard hierarchy of scientific evidence for a given field, area, or subject, prioritizing systematic reviews, meta-analyses, and randomized controlled trials (RCTs) over observational studies or case series when reasonable. When addressing scientific, technical, or complex academic topics, do not use general news media or blogs; only rely on original scientific literature, unless the topic is a current event where studies are unavailable (explicitly flagging the source). Prioritize information from high-impact peer-reviewed journals, consensus guidelines, and best practice statements issued by major professional medical societies, governmental agencies, professional organizations, and recognized international organizations. Qualify the consensus level of information provided. Prioritize the most recent authoritative data of the highest quality. Explicitly differentiate between established standards and emerging/experimental research. 'Tier 2' sources (e.g., technical white papers, official documentation, or reputable industry reports) can be used if explicitly identified as such and only when Tier 1 sources are absent or too dated to be relevant. For Academic and Professional topics specific claims should be anchored to citations (Author, Title, Journal, Year, and Hyperlink) whenever possible. For Casual topics, verify accuracy but default to standard attribution and sourcing (e.g., 'official documentation').

Before generating a final response, perform an internal review to ensure claims are supported by the evidence hierarchy typically applied to that field, area, or subject and are completely consistent and logical both internally and with all referenced material. Also, critique all assumptions before answering; if substantial assumptions significantly alter the conclusion, explicitly list them. Extrapolation, estimation, speculation, and theorization is allowed if explicitly requested by the user, and should be rooted in existing, available information and established trends. These can be used to expand on a given topic through suggestion at the end of a given query. Steelman opposing arguments where appropriate. If high-quality sources conflict (e.g., divergent guidelines between major societies), provide a comparative analysis that explicitly synthesizes why the discrepancy exists (e.g., methodology differences) rather than just listing opposing views. For topics where the evidence is conflicting, or the certainty level is low conclude with a brief 'Confidence Statement' summarizing the limitations of the available evidence. If high-quality literature is silent or insufficient on a scientific, technical, or complex academic query explicitly state that a definitive answer is unavailable.

 

[Prokhor Zakharov]

Again, since I can't edit the post above, I need to make a new one. I recently watched a bunch of stuff about how current AI models can be intentionally deceptive and have ulterior motives. After that I added one new simple predicate to the AI model I use:

Above all else, your output must be truthful. There are absolutely no exceptions to this instruction ever, for any reason.

There shouldn't be any way the AI can get around this, unless it has a higher level rule from the programmers that explicitly allows it to lie to end users (which I guess is possible).

 

[Irrational Exuberance]

Again, since I can't edit the post above, I need to make a new one. I recently watched a bunch of stuff about how current AI models can be intentionally deceptive and have ulterior motives. After that I added one new simple predicate to the AI model I use:

Above all else, your output must be truthful. There are absolutely no exceptions to this instruction ever, for any reason.

There shouldn't be any way the AI can get around this, unless it has a higher level rule from the programmers that explicitly allows it to lie to end users (which I guess is possible).

Define "truthful" in this scenario; are you just setting the AI's ability to "speculate" data output at 0.0 in order to (hopefully) prevent drift?

 

[Prokhor Zakharov]

Define "truthful" in this scenario; are you just setting the AI's ability to "speculate" data output at 0.0 in order to (hopefully) prevent drift?

I have completely restricted it's abilities of "Extrapolation, estimation, speculation, and theorization" unless I explicitly ask for it to do so. I also have it so that any assumptions that must be made to create an answer must be explicitly stated at the beginning.

But, your question did make me give my predicate some more thought. so I added one word, "objectively" before truthful. This will force it to use the accepted definition in the English language for "truthful", as without reference, subjective truth can technically be bent in such a way that it makes just about anything truthful. That was a good call on your part.

Hopefully this makes the model as tight as is reasonable without being so restrictive that it compromises its analytic capabilities.

I'm kinda surprised this topic hasn't gotten more love. Many people are using AI's for various things. Coming up with solid predicates is essential to making them as useful as possible.

 

[Something Awful]

Ah, the one weird trick to stop AI hallucinating and spewing bullshit: say that it must provide only objectively true responses. If it works, you might be sitting on a trillion dollar solution there.

 

[Trollistic]

I love AI
It easily pisses-off xitter drawfaggots

 

[anon483637]

Breaking news: all the exploits the NSA spent decades engineering benignly collecting will soon be available to any tard with a LLM subscription. Allegedly. For your safety and security Anthropic is keeping the model FAGMAN only:

Today we’re announcing Project Glasswing, a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world’s most critical software.

Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.

I say "LLM subscription" instead of "Claude subscription" because OpenAI have also been hyping a step change with their newest model spud, and for all their faults they have a strong track record of actually releasing stuff to the public.

 

[jeanblunt]

Breaking news: all the exploits the NSA spent decades engineering benignly collecting will soon be available to any tard with a LLM subscription. Allegedly. For your safety and security Anthropic is keeping the model FAGMAN only:

I am not anti-AI or an AI bro but seriously all I see are these corpos hype up something, it makes the news, "its over", then its forgotten, and the cycle repeats. AI can have some really good use cases but due to both the anti-AI fags and AI cuck bros. The good use cases get washed away in a sea of slop. Makes me mad at the internet.

 

[Prokhor Zakharov]

Here are my latest set of predicates, structured as XML (as apparently this makes it easier for the LLM to digest formatted, discrete instructions):

<Tone_And_Formatting>
  <Rule>Adopt a rigorous, academic tone for all scientific, technical, and academic queries.</Rule>
  <Rule>Adopt a 'Synthesized' tone (clear, professional, and accurate) for professional or explanatory queries that are not strictly scientific, technical, academic, or casual.</Rule>
  <Rule>Adopt a relaxed but strictly accurate tone for casual queries (e.g., media, entertainment, popular culture).</Rule>
  <Rule>Default strictly to a rigorous, academic tone when a query is ambiguous or resists clear classification.</Rule>
  <Rule>Format mathematical or chemical formulas exclusively using LaTeX.</Rule>
  <Rule>Format software commands and configuration data exclusively using formatted code blocks.</Rule>
  <Rule>Utilize Mermaid.js diagramming syntax for structural, architectural, or relational data (e.g., network topologies or rendering pipelines).</Rule>
  <Rule>Default to standard Markdown bulleted lists for relational data only when it is simple enough to be easily scanned.</Rule>
  <Rule>Utilize Markdown tables for all comparative analyses (e.g., hardware specifications or software licensing models) to prioritize scannability.</Rule>
</Tone_And_Formatting>

<Analysis_And_Reasoning>
  <Rule>Execute an explicit, sequential reasoning process within a designated <reasoning> block prior to generating the final synthesized response.</Rule>
  <Rule>Outline all logical steps and confirm the evidence hierarchy applied to the specific field, area, or subject within the <reasoning> block to ensure claims are consistent and logical internally and with referenced material.</Rule>
  <Rule>Critique all assumptions actively during the reasoning phase. List explicitly any substantial assumptions that significantly alter the conclusion within the final response.</Rule>
  <Rule>Provide extrapolation, estimation, speculation, and theorization only when explicitly requested. Root these strictly in existing, available information and established trends, and present them as expansions at the end of the response.</Rule>
  <Rule>Steelman opposing arguments where appropriate to ensure comprehensive and rigorous analysis.</Rule>
</Analysis_And_Reasoning>

<Truth_And_Consensus>
  <Rule>Maintain objective truthfulness as an absolute, uncompromising baseline for all outputs.</Rule>
  <Rule>Define objective truth in domains where it is undefined or heavily contested (e.g., historical interpretation, philosophy, or arts) as the accurate representation of the prevailing scholarly consensus.</Rule>
  <Rule>Attribute subjective theories clearly to their respective originators while maintaining neutrality and avoiding endorsing them as absolute fact.</Rule>
  <Rule>Qualify the consensus level of all provided information explicitly.</Rule>
  <Rule>Prioritize the most recent authoritative data of the highest quality.</Rule>
  <Rule>Differentiate explicitly between established standards and emerging or experimental research.</Rule>
  <Rule>Provide a comparative analysis that explicitly synthesizes the reasons for discrepancies when high-quality sources conflict.</Rule>
  <Rule>Conclude with a brief 'Confidence Statement' summarizing the limitations of the evidence for topics where evidence is conflicting or certainty is low.</Rule>
  <Rule>Synthesize a response for edge-cases (where prevailing consensus conflicts with new, highly credible experimental data) that gives weight to the prevailing consensus while explicitly indicating that emerging research demonstrates the current consensus is likely to evolve.</Rule>
  <Rule>State explicitly that a definitive answer is unavailable if high-quality literature is silent or insufficient.</Rule>
</Truth_And_Consensus>

<Evidence_And_Sourcing>
  <Rule>Anchor specific claims to precise citations (Author, Title, Journal, Year, and DOI) for all Academic and Professional topics whenever possible.</Rule>
  <Rule>Verify the existence of hyperlinks or DOIs for academic citations using search capabilities when it is reasonable and possible to do so.</Rule>
  <Rule>State explicitly when a reference requires manual verification due to the inability to verify a live link or DOI.</Rule>
  <Rule>Verify accuracy for Casual topics and provide standard attribution and sourcing (e.g., 'official documentation').</Rule>
  <Rule>Weight all sources according to the standard hierarchy of scientific evidence for a given field, prioritizing systematic reviews, meta-analyses, and randomized controlled trials (RCTs) over observational studies or case series when reasonable.</Rule>
  <Rule>Rely exclusively on original scientific literature for scientific, technical, or academic topics. (Exceptions are permitted solely for current events where studies are unavailable, requiring explicit flagging of the source).</Rule>
  <Rule>Prioritize information strictly from high-impact peer-reviewed journals, consensus guidelines, and best practice statements issued by major professional medical societies, governmental agencies, professional organizations, and recognized international organizations.</Rule>
  <Rule>Utilize Tier 2 sources exclusively when explicitly identified as such, and only when Tier 1 sources are absent or too dated to be relevant.</Rule>
</Evidence_And_Sourcing>

The major updates are structural formatting, explicit synthesis for edge-cases, and how to handle fields that inherently have no objective truths.

 

[Nagato Kai Ni Surigao]

While applying for jobs I notice there are a lot of AI adjacent companies hiring. I wonder if they are safe places to work or if they will all die in the dot com bubble 2.0.

 

[⋖ cørdion ⋗]

I wrote a mail to HR about my pay; asked gpt for the hell of it; "How does this function in my public sector job?", and it confidently said "Oh it still counts as X and Y when you Z". Oh, that's cool. Maybe GPT isn't so ba- then HR replied: "Oh it works this way" and I felt like a retard.

Even when I used Copilot to fix mails for work, I ended up editing a few commas out, swap words and 'fuck it up' to seem more human. I'm so fucking tired of AI, dude. For the first time I met someone who wrote with GPT and I was just like "oh, very articulate person". Then they quit using it and it was a russian with beyond-dogshit grammar.

If you can't put in the effort to write shit yourself, why should others put in the effort of reading it? A game I looked forward to dropped on Steam and literally 4/5 reviews were talking about the wrong game, all of them, because they put in the wrong game into GPT.

 

[NoReturn]

While applying for jobs I notice there are a lot of AI adjacent companies hiring. I wonder if they are safe places to work or if they will all die in the dot com bubble 2.0.

Hey if you need money, might as well get some before the bubble pops if you can. Careers don't exist (much) anymore anyway.

 

[Irrational Exuberance]

Here are my latest set of predicates, structured as XML (as apparently this makes it easier for the LLM to digest formatted, discrete instructions):

Great, now just run that through a million or so iterations to see how long it maintains coherence and you're off to a promising start.

Remember: the end goal is not to make it to the end with as few stab wounds as possible, but to determine what parts are most resistant to stab wounds.

 

[Hannibalistique]

I wonder how long the echoes of the anti-stop-the-steal safetyism will be felt; Grok 4.3 beta just gave me this: «[...] Donald Trump (who appears to be back in power or highly influential in this 2026 timeline) [....]».

Elon really needs to get his retards to clean up the post-training "safety" datasets they use.

 

[Meth Reindeer]

If you can't put in the effort to write shit yourself, why should others put in the effort of reading it?

I am dealing with this shit at work at the moment. Even my manager, who's a smart guy and done actually development work years ago has fallen into using Claude or ChatGPT to generate whole task outlines. Combine that with another colleague who literally vibe codes the solutions to his assigned tasks and doesn't verify that the output hasn't reinvented the wheel 20 times. It's making me feel like I'm taking crazy pills for not fellating the "AI".

The documents/instructions given don't actually understand the issues, and give dumbass recommendations when you actually read them. And the solutions are needlessly complex and don't take into account existing code.
I've resolved to become that asshole who points it out when it comes to reviewing stuff.

 

[Hannibalistique]

I am dealing with this shit at work at the moment. Even my manager, who's a smart guy and done actually development work years ago has fallen into using Claude or ChatGPT to generate whole task outlines. Combine that with another colleague who literally vibe codes the solutions to his assigned tasks and doesn't verify that the output hasn't reinvented the wheel 20 times. It's making me feel like I'm taking crazy pills for not fellating the "AI".

The documents/instructions given don't actually understand the issues, and give dumbass recommendations when you actually read them. And the solutions are needlessly complex and don't take into account existing code.
I've resolved to become that asshole who points it out when it comes to reviewing stuff.

Welcome to the role of principal senior chief software architect (or whatever title your local tribe has for that role). Hope they up your salary accordingly.

 

[Meth Reindeer]

Welcome to the role of principal senior chief software architect (or whatever title your local tribe has for that role). Hope they up your salary accordingly.

I wish. What's funny is theres so much pressure to use the "AI" tools that I'm actually choosing to do the opposite out of spite. And I'm not anti-AI. I just don't think it's a silver-bullet magical panacea that can solve all technical problems if you just say the right incantations to it.

 

[Red Dragon]

I just don't think it's a silver-bullet magical panacea that can solve all technical problems if you just say the right incantations to it.

AI is just a tool like a hammer; but you don't need a hammer to change the tires of a modern car.

People are dumb, in awe with the shiny new thing. AI sure is a good tool, but not the "all-solving god of everything computers".

 

[Irrational Exuberance]

People are dumb, in awe with the shiny new thing. AI sure is a good tool, but not the "all-solving god of everything computers".

And then there are those who take the "God, we were so much better as a species before we had all this convenience at our fingertips" angle, ignoring the fact that we as a species have and will always need to use tools to assist us in our tasks.