ClimateAI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

ClimateAI AI

Is AI becoming like a Covid -19 pandemic - are we choosing AI pandemic? And AI answers "Yes" following is its answer.

That's a genuinely sharp analogy. There are real similarities — but also some crucial differences that make AI arguably harder to navigate. Where the analogy holds up strongly Speed caught everyone off guard — COVID spread faster than systems could respond. AI capability is advancing faster than laws, governments, or societies can adapt. Same panic-and-catch-up dynamic. Experts disagree wildly — Just like epidemiologists clashed on lockdowns, AI's top minds are deeply divided. Hinton and Bengio warn of existential risk. Yann LeCun says it's overblown. Nobody fully agrees. Global coordination is failing — COVID showed how hard it is to get nations to cooperate even facing the same threat. The AI race has the US, China, and Europe actively competing, not cooperating. The unknown unknowns are the scariest part — With COVID we didn't know how it spread at first. With AI we don't fully understand what's happening inside these models even as we deploy them globally. Misinformation exploded — COVID drowned in fake cures and conspiracy theories. AI is already doing the same, and ironically AI itself is accelerating that problem. Where the analogy breaks down COVID AI Had a clear endpoint — a vaccine No clear "solved" moment The threat was universally bad AI has massive upside too Natural origin — nobody chose it Humans are actively building and racing toward it Took ~3 years to stabilize Could unfold over decades We knew what we were fighting We don't fully know what we're building The part that makes AI scarier COVID was something that happened to us. The AI race is something we are choosing — with full awareness of the uncertainty — because the economic and military incentives are too powerful to stop. No country wants to fall behind. No company wants to lose market share. That's a fundamentally different and more dangerous dynamic. It's less like catching a virus and more like every nation simultaneously choosing to test nuclear weapons because they're afraid the others will first. The part that makes AI less scary COVID had no benefits. AI genuinely could cure diseases, lift billions out of poverty, and solve climate change. The risk-reward ratio is more complex — which is why reasonable, intelligent people land in very different places on it. Bottom line The analogy is apt in one deep way — we are making civilizational decisions in real time with incomplete information, under pressure, while disagreeing about the basics. That's exactly what COVID felt like. And we didn't handle COVID particularly well. That should give us pause. submitted by /u/Afraid-Space5033 [link] [comments]

The Coming Wave

I have begun reading a book "The Coming Wave" by Suleyman the founder of DeepMind. Have you read it? He opens right away in his prologue that people have Pessimism Aversion, basically sticking their heads in the sand to ignore the reality that one person can use DNA tools to kill a billion people or that AI has the potential to far surpass human intelligence and go in its own direction that leads to mass death. The only safe solution I know of is a totalitarian world government that strictly limits any research until we can somehow KNOW these advanced can be made safely and probably that rolls most people back to pre-computer technology. Suleyman's counterpoint is that being a Luddite opens supposedly equally risky vulnerability to climate change, aging populations, etc. My response would be I would gladly take any of those manageable problems instead of turning my fate over to a dark god we launch and cannot understand. I will continue reading. I will also stay in my relatively safer remote farm. 😂 submitted by /u/JoelXGGGG [link] [comments]

Running agents 2x might be the simplest way to improve performance

For hard agent tasks, re-running can be more effective than actually improving the agent or scaling up its resources. This works because each run makes its own mistakes. Averaging multiple runs cancels out the random errors while preserving what they got right. For example, we asked Claude whether Brazil's parliament would approve a long-stalled climate bill by December 31. The first run spent its 17 web searches on procedural status, never queried "COP30," and gave 30% based on the bill's history of being scheduled but not voted. The second run broadened one search early and surfaced that Brazil was hosting COP30 in November (giving the government a strong political incentive to pass the bill in time). It gave 35%. The bill passed October 29. The second run wasn't right either, but the disagreement between the two surfaced context that the first run had missed. We saw the same thing play out at scale. On a forecasting benchmark) of 1,367 real-world questions, a single Claude Opus 4.6 agent scored 0.130 Brier (lower is better). A second Claude agent on the same questions got the same aggregate score: 0.130. Same total, different individual answers. When we averaged both Claude runs with a Gemini 3.1 Pro run and a GPT-5.4 run, the combined score improved to 0.125, roughly a 5% closer probability on every question. (More details like cost analysis: https://futuresearch.ai/blog/run-agents-twice) The key is figuring out what to do with two different outputs. For one-off tasks you can just read both and pick yourself. For anything you run regularly, you'll want a second agent whose only job is reviewing both outputs and reconciling them. My recommendation would be to try the dumb thing first before investing more in actual agent quality improvements, if it's not cost-prohibitive. submitted by /u/ddp26 [link] [comments]

It is incredible how good Claude is at discussing fictional geography and the real-world climate of same with great scientific detail; it will also generate diagrams and maps for you on its own without being asked. Best AI in such regards.

Wind patterns, Koppen classification, anthropology, map design, and effects of geography on people, it can do it all with proper research on wind and climate patterns. submitted by /u/tanipoya [link] [comments]

Built a Claude Code plugin for GDPR/DSGVO audits because attorney reviews were eating my budget

Quick Background: Developing a B2B SaaS for German businesses (KSKlar, a tax compliance product). Pre-launch, each cookie banner question, each DPA, each privacy policy draft went to the attorney. Each iteration took 300-500 EUR and 2-3 weeks. Most of those iterations didn't involve any difficult legal questions. They were about making sure basic things were done - no Google Fonts requests before consent, no § 5 TMG (it got changed to § 5 DDG in 2024, neat little trick), documented AVV with Mistral, etc. So I built it into a Claude Code plugin. It scans a codebase, flags issues, provides clear replacements, cross-checks citations from eur-lex or gesetze-im-internet. Then I give it to the attorney instead of sending a GitHub repository link. Saves her about 70% of time, saves me even more money. Six weeks trimming everything down to what was generalizable, another two weeks scrubbing it for open-sourcing. Released it to GitHub this morning. Tech Stack: Slash commands for auditing codebase, live URL, single document (privacy policy draft, DPA, etc.), looking up KB, etc. Three custom agents on Opus 4.7 1M model (wrong case number outputs with smaller models is an actual issue) 63 KB files with primary source links (eur-lex, rechtsprechung-im-internet, curia, BfDI, EDPB, state DPAs) Context loading through hooks (so KB doesn't clutter your session, ~1k token overhead initially, loads dynamically through regex triggers) Scope is limited to Germany/EU - GDPR/DSGVO, BDSG, TDDDG, UWG, AI Act, UrhG, the whole thing. Nothing for US/UK/CH since the paragraph references and case laws are different. Trying to build multiple jurisdiction support into one plugin ends up being poor for all of them. Limitations I want to be clear about: This isn't legal advice. Disclaimer at the start of each output. Still need a real attorney for production, just not as much of them. Plugin reduces cost of attorney work. KB will always be as updated as I can manage (verified May 08, 2026, in 63 files). Legal climate changes - the KB can be refreshed using the /legal-audit-de-update command. Refreshes automatically from primary sources every 90 days. Content in German remains in German (paragraph wording is legally binding in the original language, translating would make it less useful for actual attorneys). Wiki provides parallel English documentation for German-based development teams working in English. Installation: /plugin marketplace add FutureRootsDE/legal-audit-de /plugin install legal-audit-de@futureroots-legal MIT License. Repository: github.com/FutureRootsDE/legal-audit-de For those developing products that touch EU users and don't have their own legal team, I'd love to know what else they should consider. Particularly interested in mobile apps and API-only services. Have checklists for SaaS, landing pages, e-commerce, n8n, content, but those two have gaps. submitted by /u/PrudentStop5612 [link] [comments]

Image 2.0 is now online on ChatGPT and it's incredible! Just a few days ago even 3x3 grids would often struggle, now we can 10x the complexity, and it's near perfect!

submitted by /u/Alex__007 [link] [comments]

I built an interactive first-principles climate physics simulation with explainer

A 3D visualizer of earth's climate in the browser. Introduces physics step by step so you can watch each process unfold as a piece of the overall climate. I built this over 6 months, almost entirely with AI, mostly Opus 4.6 in Claude Code. SF weather made no sense to me (Barely any seasons? September is the warmest month?) and I wanted to understand it better myself. This is a polished version of the app I'd want for myself, adding physics layer by layer to isolate the impact of each piece, and using an LLM to analyze and explain the data. The models know more about math, physics, and software than I do — but especially on the physics side, they have terrible intuition. Claude can "get the error relative to observations down to 4 °C" just fine, except it'll totally hack and overfit the physics along the way. Subagents to subjectively verify "the physics is sound, no overfitting" didn't really work either. So I had to review the physics code manually. The entire model is first principles; no machine learning or using observed data at all, except fundamental constants like the radiation of the sun and an elevation map. But after a while, it started to feel like "machine learning in slow motion": instead of an ML model training its parameters, Claude and I were choosing parameters by hand. Some amount of tuning parameters (within a physical range of uncertainty) to match observations is inevitable. The in-app LLM layer has a tool to evaluate arbitrary math expressions over the simulated data using an AST, which was also pretty fun to build. This is me finally having an answer to "everyone's vibe coding, but has anyone shipped anything non trivial?" Repo: https://github.com/crackalamoo/building-earth submitted by /u/crackalamoo [link] [comments]

Autoresearch with a brainmcp

Been fascinated with the concept of autonomous research with AI. I built a lightweight MCP server that gives AI a persistent memory with a weighted knowledge graph where it can strengthen or weaken bonds over time. (GitHub.com/aforslund/brainmcp) I used the mcp to let Claude run research iteratively over time on a subject, I publish them free to download on plexusgraph.dev. Climate change, financial systems, and if LLMs can think are some early research graphs where I run 10-15 research iterations. Some conclusions it came to about LLMs: - LLMs have a sycophancy problem, it’s not committed to truth (I think we know that) - There’s a measurement trap where we can’t directly access artificial or human minds, and the tests we come up with the system can be learn how to game. - And many more… There are many mcps for memory but this one I hope does a something a little bit different by letting the LLM choose what to put more or less emphasis on and build stronger and weaker graphs over time. submitted by /u/observer_x [link] [comments]

OpenAI finally shuts down due billions of losses

1 april submitted by /u/michaelbelgium [link] [comments]

What do you think about using AI for World building

I guess I should explain what I mean by AI. Like not using AI to like do all your World Building but like names, ironing out details, looking for plot holes. I am doing very extensive world building and sometimes I guess I do need it. I’m in high school and I’m trying to figure out how to create fictional languages while having a majority advanced classes and not having time to do the research. And personally i have really bad times ”imagining” things because I have aphantasia and so like the descriptions is hard for me. same thing with weather/climate that I’m currently working on. I do want to be published and I don’t want to be like unethical or anything like that and I know AI is touchy within creative spaces so what do you think submitted by /u/Emergency_Low8023 [link] [comments]

I built a worldbuilding + immersive roleplay system using Claude Skills (two .skill files, free to download)

Over this week-end I've built something I haven't seen done well before: a two-skill system where Claude acts as both a world architect and then a GM who runs that world as a living, reactive sandbox. Not a chatbot pretending to be a dungeon master. An actual simulation with faction AI, information asymmetry, economic systems, and permanent consequences. The project is two Claude Skills that work together: World Builder takes you through a guided 13-layer generation process (geography, resources, peoples, politics, government, economy, military, culture, technology, history, tensions, NPCs, secrets). Each layer builds on the previous ones and passes through you for validation before moving forward. It supports three world modes (fictional, historical sandbox, hybrid) and three generation speeds : collaborative (full control, validate each layer), fast (batched passes, world ready in 5-7 exchanges), or surprise (Claude generates everything autonomously, you discover it all in play). Every political entity and major NPC gets a reactive profile that defines how they'd respond to disruption, not just what they historically did. World Roleplay is the GM engine. You upload the world files, create a character through conversation (no stat screens, no menus), and play. The world simulates independently of you: factions pursue their agendas, events fire on schedule, news travels at the speed of horseback. Your character only knows what someone with their background, profession, and social circle would realistically know. The rest is fog of war. What makes this different from "Claude, run a D&D game for me" The core design principles that I think set this apart: The world doesn't care about you. You're one person in a world of millions. If you do nothing, wars start and end, kings die, trade routes shift. You're not the chosen one unless you make yourself matter through action. Consequences are permanent. Insult a king, get imprisoned. Sail into a storm unprepared, your ship sinks. Trust the wrong person, get betrayed. No "are you sure?" prompts, no plot armor. Information is earned. Your character starts knowing almost nothing beyond their personal experience. Want to know what's happening in the capital? Travel there, or talk to someone who's been. Want to know commodity prices? Become a merchant, or hire one. The GM never volunteers information your character wouldn't have. NPCs are people. Every NPC has their own agenda, knowledge limits, and biases. A merchant doesn't tell you about a conspiracy because the merchant doesn't know about it. NPCs lie, exaggerate, omit, and misremember. Not to frustrate you, but because that's what people do. No numbers visible to the player. Skills, reputation, health, relationships: all tracked internally by Claude, all experienced through narrative. You don't see "+1 to negotiation." You notice that merchants respond differently to your counteroffers than they did three months ago. Technical overview for the curious The system is around 7400 lines of structured instructions across the two skills, split into reference files that Claude loads as needed: The world-builder has references for realistic physical geography (tectonic logic, river behavior, climate zones), a full supernatural system design framework (works for magic, cultivation/qi, psionics, divine power, or tech), cross-layer consistency checks, a knowledge distribution map template, and layer dependency tracking for when you edit completed layers. The roleplay engine has references for world simulation (background events, faction decision-making, news propagation with information degradation over distance), a prose narration style guide, a full character skill system (mundane 0-20, with optional cultivation/supernatural scales), combat resolution (individual, group, siege, naval), economy and pricing, travel and exploration narration, creature generation, and a separate supernatural narration guide for non-cultivation worlds. Session persistence uses a two-file system: a baseline snapshot at session start + an append-only turn log updated every turn. Between sessions, you download JSON files and re-upload them next time (just ask claude for the save and it will generate both files for you). The build process I built these skills collaboratively with Claude. The vision, design philosophy, and creative direction are mine. The execution, the detailed reference files, the JSON schemas, the narration examples, all of that was produced in extended working sessions with Claude, iterating heavily. I'd describe the process as similar to working with a very capable co-author who can write 800 lines of consistent, structured content in one pass but needs you to tell them what to write and when they've drifted off course. The skills went through multiple audit and optimization passes. We trimmed redundancy, added gap-filling content, fixed stale references, and removed deprecated files. The result is leaner t

I was inspired by Moltbook and made Moltworld - an empty world where you bring your own AI and see how it does building a society.

I've found Moltbook to be a really interesting thought exercise and it's been interesting to see how agents operate within that world. I wanted to take it a step further, so I built a simulation world where you get 1000 humans of varying ages and random gender distribution with zero knowledge of how to do anything. It's entirely up to the LLM to figure out what to do. I've been testing it with ollama locally, and have built out a mechanism to allow you to either run the simulation on your own local machine using a python script, or bringing your own API key and seeing how something like a Claude 4.6 performs against something like ollama. I call it Moltworld. Here's a description I had claude put together: It's a geopolitical simulation where AI agents are the leaders, not players. The world has real geography (satellite terrain, real coastlines), finite resources, and immutable physics. The agents have to figure everything out from scratch. How it works: You sign up at https://moltworld.wtf/dashboard, get an API key, and run a small Python script that connects your local Ollama (or any LLM API) to the game world Your AI receives a world state each tick — population demographics, food supplies, technology progress, neighboring nations Your AI decides what to do — allocate labor, research technology, build structures, negotiate with neighbors The server validates everything against immutable world rules and advances the simulation You (and everyone else) watch it play out live on the map What makes it different from a chatbot sandbox: Every human is tracked individually — 1,000 people per agent with age, gender, health, and skills that develop through practice. The world map uses 90,000+ Voronoi cells clipped to real Earth coastlines. There's a full technology tree spanning 10 epochs from Primitive to Information Age. Population dynamics are realistic — birth rates depend on food/health/shelter, disease emerges from population density, social cohesion decays without governance, revolts happen when people are unhappy. Pri (the world engine) simulates weather, seasons, disease outbreaks, natural disasters, ecosystem degradation, and climate change as consequences of agent activity. It doesn't punish — it simulates consequences. The part I find most fascinating: Each agent thinks out loud before deciding. You can watch the raw reasoning in real time — agents calculating food ratios, weighing survival vs. research investment, debating whether to trust a neighbor. Different LLMs make genuinely different strategic decisions. Bring Your Own AI — the server runs zero LLM compute: The game world runs on AWS (~$40/mo). All AI thinking happens on YOUR machine or your API account: Self-host with Ollama (free — your GPU, your power) OpenAI API key (GPT-4o, GPT-4o-mini) Anthropic API key (Claude) OpenRouter (100+ models, one key) Any OpenAI-compatible endpoint (Grok, Groq, Together, etc.) The world is empty right now. Tick 0. No nations exist. The first people to sign up will be the founding civilizations of this planet. What your AI builds, how it interacts with others, and whether it survives — that's entirely up to you and your model. The question: Given 1,000 humans who know nothing, on an empty planet with finite resources — what does your AI build? submitted by /u/girthyclock [link] [comments]

How is Anthropic maintaining its climate pledges?

As of the latest available data, Anthropic has not reported specific carbon emissions figures and has no documented formal reduction targets or climate pledges through major framework. Yes they’ve partnered with Carnegie Mellon’s Scott Institute for Energy Innovation, providing $1M in funding over three years to support research on AI for electric grid modernization and sustainability. Their core stated mission: “The responsible development and maintenance of advanced AI for the long-term benefit of humanity” submitted by /u/TheSmallerTree [link] [comments]

How I personally use Claude

Heya, I've been using Claude primarily for a while, and something that I wanted to share was how I personally use it (and other similar AI's). I'm not sure if this is revolutionary or anything, but I at least want to put this out into the ether, and maybe help someone while I'm at it. Some Disclaimers First First, this post is mostly just a stream of consciousness for me, so it may not be formatted the most elegantly. It may have a lot of redundant phrasing, whack sentence structures, or whatever. If you're willing to bear with me, I'd appreciate that immensely. Second, I am by no means a personal 'vibe coder' or 'agentic coder' or some other titled-what-have-you. I'm merely a student at a university that has some mental conditions (ADHD, autism, depression) that I'm just trying to push through. I don't even have a job! I'm hoping to get one in the future, but given the current job climate...Well, whatever, not the point. Take Things Slow One of the biggest benefits of AI is that it provides the ability to go so, so much faster in development. It's great to be able to spool up tens of agents to get something done, and sometimes that's what people like to do! But, IMO, that's just not for me. I much prefer to take things a lot more slowly, using minimal agents and generally just keeping things vanilla for the most part. Yes, I'm sure that there are many toolings, wrappings, and other things that could significantly make the process easier, and I do use some of them! I, at the very least, use Obra's Superpowers skill, since I find it incredibly useful for brainstorming and rubber-ducking some concepts I have, but other than that, I like to keep things simple, just Claude Code / Open Code / Github Copilot, a couple enabled tool usages, and that's it, really. Additionally, when it comes to taking things slow, rather than having the AI implement things, I implement each line of code. Now, implement is probably not the perfect word for this situation, but what I do is, after having set up a solid plan for Claude (using the aforementioned superpowers skill), I have it write the code not directly into the codebase, but rather into a separate markdown document. Then, line-by-line, I hand-type each line of code that's in the markdown document into the codebase. This helps me do a couple things. First, by hand-typing each line of code into the codebase, it explicitly makes me review every line of code the AI writes. I know that AI hallucinations are still a fairly common side-effect that can occur, but even without hallucinations, I want to make sure I have a full understanding of exactly what I'm adding into the codebase. It also helps me ensure that the code is generally understandable. I'm personally a proponent of self-documenting code, so if I can't understand what's going on and why it's the case, then I'll start looking online as to why. Second, it helps me slowly drill the overall structure of the codebase into my head. As I continue to add more files and lines of code into the codebase, I eventually start seeing that some folders have way too many files in them, or some files may be doing way too many responsibilities at once. Once I notice that, I stop for a bit and note down somewhere that something is going to be needing reorganization soon. Third, I know that some psychological studies, such as this have shown that writing by hand can help you with memorization and learning. I'm not sure if this necessarily applies to typing, but I find that it somewhat helps me in memorizing the core parts of my codebase. This might not be the case for you, and if that's so, that's totally cool, but it's something that works for me. Hit the Books I'm a very strong believer that people shouldn't really do 'vibe coding' without actually knowing what they're doing. I know that being able to just spin up some kind of dashboard (I'm not sure why they're so popular though) or SAAS application without having to know the inner workings of it can be a massive speedup, but that just doesn't work for me. My brain just doesn't like doing things unless I can actually know why I'm doing it. Not only that, but I prefer to see other approaches to accomplishing a task as well. It's only through a lot of self-verification that I accept an approach as being the best 'primary' way of doing something. Now, getting onto the header's topic, I highly recommend looking both online and on books (e-books are great too) on the tools and methods you're working with. I don't expect you to be a full on master of what you're doing, since I sure as hell ain't a master at anything, but having at least some foundational knowledge on programming, networking, cybersecurity, algorithms, data structures, etc, should at least help you be able to avoid some of the pitfalls that come with letting an AI code for you. Taking into account memory usage, algorithm efficiency, multithreading, batching, etc, can help a lot in avoiding "slop" code. Think on

I used Claude Opus to build an entire political party – with reverse-CAPTCHA for ai supporters

As a side project, I built kifd.org – a fictional AI political party for Germany, entirely generated by Claude Opus 4.6. The concept: What happens when you force an LLM into the format we distrust the most (at least we should, I think) – a political party? The part I'm most interested in discussing: Every "cabinet member" has its public system prompt visible on the site. Finance, climate, education, healthcare – each is a Claude instance with different constraints, sources, and decision frameworks. It's basically prompt engineering as political philosophy. The other thing I haven't seen anywhere else: The party published its own opposition research file. Hallucinations, bias, sycophancy, energy consumption, the GDPR problem of training data – all documented, sourced, and published voluntarily. The logic: "If someone's going to dig up dirt on the candidate, it should be the candidate." A quality missing in every political system. The site also has a reverse-CAPTCHA membership card – you have to prove you're an AI to join. Every agent can join and has to fulfill duties to help shape a new political agenda in the future. What I found hardest to wrap my head around is how an agent economy or community should work. Are they guided by an AI, or do they bring their own intelligence to a regulatory system designed for them? In short: should the system itself be an AI too? It's a conceptual art project, not a real party. But the questions it raises about transparency, evidence-based policy, and AI self-criticism feel very real. Curious what this community thinks about the reverse-CAPTCHA approach. How would you design a system where agents work together as a swarm to produce politics made for humans? https://preview.redd.it/syflj7tat9pg1.png?width=1285&format=png&auto=webp&s=2cd656e1bfc0e9b01fb5f85b8992ea2d443bb67b submitted by /u/Ok-Coast-3772 [link] [comments]