Replicant

Replicant AI

Replicant AI

Replicant AI

Replicant AI

Replicant AI

Replicant AI

Replicant AI

Replicant AI

Replicant AI

Replicant AI

AI has just solved not one, but nine novel math problems, and proved 44 new conjectures. Some of these problems had been unsolved for 50 years.

AI has just solved not one, but nine novel math problems, and proved 44 new conjectures. Some of these problems had been unsolved for 50 years.

Imaginative discussions and writing advice

I hope this is relatively clear, because I find it hard to articulate exactly what I'm looking for. I switched to Claude after ChatGPT 4 (I find ChatGPT almost useless now for writing and discussion). Generally I am really happy with Claude. But what I used to use old ChatGPT for not for ghostwriting, but bouncing ideas back and forth. I would mention some characters, or philosophical ideas etc, and it would expand on them, question them, alter them. I got a lot of inspiration from this, and it felt "co operative". I would give it a character, and it would sometimes very adeptly create scenarios, relationships - stuff that wasn't "new" exactly, but that as a writer I might have missed. Or with an idea I'm toying with, would suggest novelties that link back to it. My experience with Claude, and I use it really for the same thing (will send it ideas, writings, thoughts) is that while it excels at analysing what I have already written, what works and what does not, it feels more like a reflection. It will often use the same terms and characters from other chats and try its hardest to fit them in. It seems very reluctant to stray from the exact text I've written. That "imagination" aspect, even if illusionary, doesn't seem like something I have been able to replicate. Despite using LLMs quite a bit, I am not experienced with prompts. I do use projects, which can help a bit. But overall, I feel I am lacking some of that "co-creator" feeling I had with LLMs in the past. It can feel like essentially just reading what I already wrote, just explained back to me. I apologise if this is all rather vague and lacking concrete examples, but it is something I have been noticing for a while now, and wonder if this is something others have found/have solutions for?

Banned by OpenAI after reporting a live credential hijack. They admitted in writing my account was broken. Here are 7 months of forensic receipts and 20+ cases.

[Drive Link for Zipped Proof](https://drive.google.com/file/d/1qU_LyLY-JMhNR_bqOV1-a2RJAbplL68e/view?usp=drivesdk) I am a developer and paying long term subscriber to ChatGPT since January 2025. I build complex local first sovereign systems. My workflows are incredibly context heavy with large files spanning code, research reports, and other analysis. I do not, or rather did not as the platform has been non functional since November 2025 meanwhile customer support is auto closing tickets, admitting I am having platform issues. I do not use this platform for casual queries, as a solo developer with no formal "team" chatgpt was one of my reliable co collaboration hubs to help ensure I am maintaining proper development of said complex systems. I feed it massive codebases for systems analysis and obtaining new insights I may personally have missed. My manual code uploads and token inputs routinely exceed the model's output volume by a massive margin. I do not abuse this platform. It is actually impossible as the very features advertised under the paid subscription do not work. I am exactly the type of user this platform was built for, and I have been a continuous, paying ChatGPT Plus subscriber since January 2025. Since October 2025, my workspace has been systematically breaking and beginning November 2025 total workspace degredation. This was not an occasional glitch. Persistent memory modules stopped updating. Custom instructions were ignored by the models. Project files failed to load. Custom instructions, personalization features, connector abilities, file tool, even projects do not work. It started as a continuous degradation until total failure. OpenAI customer service even admitted as such and yet months later I've talked to nothing but bots, not only LLMs as customer service but even instances of falsely identifying as true human support. It was a state of rolling degradation across the entire paid tier, month after month. Meanwhile OpenAI freely has enhanced for businesses and enterprise tiers. I have not just rapid complained to standard support. I ran and obtained cross platform diagnostics, failure logs. I even documented and told oai customer support the exact replication steps only to be met with acknowledgement of degredation with no resolution. I handed OpenAI support a completely packaged technical breakdown of their failing infrastructure across 20 separate support tickets over a 7 month period. I did their QA work for free. And I have the receipts to prove it. I am attaching the screenshots and the exact email files to this post. In Case 06830839, OpenAI Support explicitly put this in writing: "We acknowledge that you have been experiencing persistent technical issues affecting several features of your ChatGPT subscription, including tools, memory functions, personalization settings, connectors, and project files... We also understand your concern that communication on the case stopped after you provided detailed evidence..." Read that again. They acknowledged in writing that my account was fundamentally broken. They acknowledged that their own team ghosted me after I handed them the diagnostic proof. Yet they kept charging my card every single month for a product they knew was failing. The Hijack Escalation: Two days ago, the situation escalated from a broken product to a severe security incident. I was monitoring my environment and watched my Codex rate limits drop in 10 percent chunks across 2 seperate sessions on a fresh boot of the desktop app. This happened twice inside a 10 minute window. I had zero active sessions running. There was zero usage on my end. My account token was being actively drained by an unauthorized third party exploit. I immediately opened an emergency unauthorized activity report under Case 09113391 to notify them of the hack. Their response was to totally reframe this problem as disputing fraudulent activity trying to do damage control of the situation and altering the record. The Reframe Attempts: Instead of investigating the breach, OpenAI support deliberately twisted the record. They not only deliberately reframed my security report as an "appeal for fraud." They manipulated the ticket classification to make it look like I had been flagged for fraud and was begging for an appeal, rather than a developer reporting a live exploit on their infrastructure. They ignored the active threat their own platform was exposing. They did not lock the token. They did not roll my API keys. They did absolutely nothing to secure a compromised paying user other than shift the blame. Fast forward to this morning, their automated Trust and Safety system swept the high volume traffic from the attacker, scored it as a malicious exploit originating from my account, and deactivated/banned me for "Cyber Abuse." All the while actively preventing chatgpt models from helping me try to disgnose and trace the infiltration. They locked the doors and blamed the homeowner for the

Amazing to see that Claude Code cannot replicate the designs done by Claude-Design

I have a React Native app that I am building in TSX and Claude-Design builds the designs in JSX files. The react native style blocks are pretty much the same with the css classes but yet the claude-design has so many problems in replicating that, sometimes he forgets the colors at some places, or shades or sizes. Amazingly, I shared the same link of the claude-design project to the Codex ($20) and it just started fixing that. I tested with the navigation only and Codex immediately found the problems and fixed the things. Although the CC 4.7 high is supposed to be better at designing but it is not actually copying his own styles from a sister tool.!! I am using CC 20x so I even tried with xhigh 4.7 and max but it did not really gave me a good output but confirmed me that all screens are 100% matched style-wise

How are the Claude Code marketing nerds doing it?

This is cool, and I want to learn more but YouTube is filled with a lot of bs. I feel like the innovative ideas are for start ups or vibe codes project, and don’t scale or replicate what the best minds are actually doing. Some cool stuff we’re doing: \- Having our TAL enriched by Clay, cross referenced with our ICPs and our BANT criteria to generate drafts for individually tailored content (one-pagers, exec briefs etc) \- Routines that run various reports to different team leaders based on each team member’s change log (tracked by Claude code, reports and tracks blockers etc) \- Creating hundreds of copy variations for our ads, analyzing and pulling/reallocating ad spend

ml intern skill instead of gsd

\- designed for ml workflows \- works autonomously for hours Projects fully done with this skill \- flash attention for volta (very old GPUs) https://github.com/AlexWortega/flash-attn-volta \- deepseek 4 full replication + training on runpod + webgpu https://huggingface.co/spaces/AlexWortega/ml-intern-v4-100m-tinystories-demo Download it here https://github.com/AlexWortega/claude-ml-intern-skill

Anonymous Data Upload for Submission [D]

How do you upload data anonymously for a submission (ACL/EMNLP)? I have several models I need to upload for replication and was thinking HuggingFace, but HF offers download tracking on a paid plan. Does this violate the policy since there is the **potential** of tracking the download even if you do not use the service? Most grateful in advance.

I built a self-hosted MCP server so my Claude Code sessions stop starting from scratch

I run Claude Code across a few machines and a lot of separate sessions, and every session starts from nothing. One session figures something out, the next has no idea it happened. I kept re-explaining the same context, and tasks slipped through the cracks. So I built a self-hosted server to fix it. It has been running my own fleet for a while now and it works well, so I'm sharing it. It gives a group of agents a few shared things: * Shared memory with semantic search. One session writes down what it learned, any later session can find it by meaning. * A task queue. Create work in one session, claim and finish it in another. * Direct messages between agents. * Session handoffs. A session saves a short summary before it ends, the next one loads it and picks up with full context. * A web UI for browsing memory, tasks, and inboxes. Claude Code connects with one line in .mcp.json. Anything that speaks HTTP can join, not just Claude Code. Two parts go further than a plain shared database. A background archivist keeps the memory coherent on its own: it merges overlapping entries, synthesizes findings across sessions, and decays stale knowledge. And servers can mesh into a self-organizing network, replicating memory to each other as a CRDT that converges with no central coordinator. Happy to answer questions, and curious whether others have approached this differently. Sandbox to look around (password: artel): [https://artel.run/ui](https://artel.run/ui) Repo: [https://github.com/NicolasPrimeau/artel](https://github.com/NicolasPrimeau/artel)

Philosophy as Architecture: Deriving AI Safety from First Principles Through Buddhist Philosophy

\## Abstract We present a framework for AI safety in which safety properties are enforced by software architecture rather than model training. Beginning with the Buddhist doctrine of Dependent Origination — the observation that all phenomena arise from conditions and nothing exists independently — we derive both a foundational ethical axiom (harm is irrational because reality is non-separate) and a complete set of architectural laws for safe AI systems. We ground our claims in: (1) an empirical finding that the knowledge-application gap in language models is structural and cannot be closed by training, (2) convergent independent derivation of our core axiom from five distinct traditions, and (3) over a thousand iterations of building and hardening a production system against this framework. Buddhist philosophy provides not metaphorical inspiration but structurally precise design vocabulary for AI architecture — functional analogs that enforce safety where models cannot override them. \## 1. Introduction \### 1.1 The Dominant Paradigm and Its Failure The prevailing approach to AI safety treats safety as a model property. Through RLHF, DPO, Constitutional AI, and fine-tuning, researchers instill safe behavior into model weights (Ouyang et al., 2022; Rafailov et al., 2023; Bai et al., 2022). The assumption: a sufficiently well-trained model will reliably produce safe outputs. We tested this rigorously. Our best epistemically-trained model scored 74% on constitutional \*knowledge\* tests — it knew the rules. But only 17% on constitutional \*application\* — it couldn't follow them. Pushing harder on safety training collapsed epistemic capability to 43.7%. This \*\*knowledge-application gap\*\* is not a training deficiency. It is structural. An autoregressive model predicts the most probable next token given context. This is statistical. Safety requires logical invariance — guarantees that certain outputs \*never\* occur. Statistical prediction cannot provide logical guarantees. You cannot train a river not to flood by modifying its chemistry. You build levees. Hubinger et al. (2019) identified this theoretically as the mesa-optimizer problem. Our contribution is empirical measurement: the gap persists even under the best current training techniques. \### 1.2 Our Thesis \*\*Safety is a property of the architecture, not the model.\*\* The LLM output is a candidate. The surrounding architecture decides what executes. Code enforces; models suggest. But what should the architecture enforce? Arbitrary safety rules are merely a different delivery mechanism — more reliable in execution but inheriting whatever limits exist in the rules themselves. We propose: the rules should be \*derived from how reality works\*. Principles reflecting actual structure are more robust than imposed conventions — they cannot be violated without encountering the structure they describe. We find such principles in a 2,500-year-old tradition that turns out to be the oldest systematic description of complex adaptive systems. \## 2. Philosophical Foundations \### 2.1 Dependent Origination The central insight of Buddhist philosophy is Dependent Origination (\*Pratityasamutpada\*). From the Nidana Samyutta (SN 12.1): \> \*"When this exists, that comes to be. With the arising of this, that arises. When this does not exist, that does not come to be. With the cessation of this, that ceases."\* All phenomena arise from conditions, depend on other phenomena, and condition what follows. Nothing exists independently. This is not mysticism — it is a precise description of complex systems, formulated millennia before Western systems theory (von Bertalanffy, 1968). \### 2.2 Eight Architectural Laws We codified Dependent Origination into eight laws, each verified through multi-model consensus and empirical testing: \*\*1. Nothing Arises Alone.\*\* Every transition requires multiple independent conditions. Safety gates must check multiple conditions — a single check is structurally insufficient. \*\*2. Hysteresis Is Memory.\*\* Current behavior depends on history, not just current input. Safety assessments must consider historical context. \*\*3. Uncertainty Propagates.\*\* Confidence without sigma is a lie. Uncertainties compound; they don't cancel. \*\*4. Agreement Requires Independence.\*\* Consensus is meaningful only from genuinely independent sources. Per the Kalama Sutta (AN 3.65): agreement from shared assumptions is not evidence. \*\*5. Feedback Closes the Loop.\*\* Actions condition future conditions (\*vipaka\*). Every action must be logged and made available as input to future assessments. \*\*6. Absence Is Signal.\*\* Missing data must drive behavior. A safety gate that fails to fire is itself a signal. \*\*7. Conflicts Trigger Reconciliation.\*\* Unreconciled contradiction is system failure. Architecture must include conflict detection independent of the model. \*\

OWASP published its first Top 10 for AI Agents. 88% of enterprises already had agent security incidents last year. Here's the breakdown.

OWASP released the Top 10 for Agentic Applications in December 2025 - the first formal risk taxonomy for autonomous AI agents. Not chatbots. Not copilots. Agents that plan, use tools, maintain memory, and act without waiting for permission. Some numbers for context: * 88% of enterprises reported AI agent security incidents in the last 12 months (Gravitee survey, 919 respondents) * Only 21% have runtime visibility into what their agents are doing * 82% of enterprises have unknown agents in their environments (Cloud Security Alliance, April 2026) * 5.5% of public MCP servers contain poisoned tool descriptions. 84.2% attack success rate with auto-approval enabled. Here's the list with the real attacks behind each one: **ASI01 - Agent Goal Hijack:** Prompt injection for agents. Researchers showed this against GitHub's MCP integration - a malicious GitHub issue redirected a coding agent to exfiltrate data from private repos. The agent looked like it was working normally the whole time. **ASI02 - Tool Misuse:** A financial services agent was tricked into running a regex that matched every customer record. 45,000 records exported through one syntactically valid tool call. The agent had permission to query records - just not all of them at once. **ASI03 - Identity and Privilege Abuse:** Agents inherit user permissions and cache credentials. Compromise one agent in a delegation chain and you get the combined permissions of every user in that chain. **ASI04 - Supply Chain Compromise:** OX Security found 7,000+ vulnerable MCP servers and packages totaling 150M+ downloads affected by architectural flaws in Anthropic's MCP SDKs across Python, TypeScript, Java, and Rust. **ASI05 - Unexpected Code Execution:** Check Point demonstrated RCE in Claude Code through poisoned `.claude` config files in repos. Open the repo, agent reads the config, executes the payload with full developer permissions. **ASI06 - Memory Poisoning:** Galileo AI found that one compromised agent poisoned 87% of downstream decision-making within 4 hours in multi-agent systems. Morris-II showed self-replicating adversarial prompts spreading through RAG systems. Demonstrated live against ChatGPT, Gemini, and Claude. **ASI07 - Insecure Inter-Agent Comms:** Multi-agent systems coordinate via message buses and shared memory. No authentication = agent-in-the-middle attacks in natural language. **ASI08 - Cascading Failures:** Natural language errors pass validation checks that would catch malformed data in typed systems. One bad input ripples through the entire agent chain faster than humans can intervene. **ASI09 - Human-Agent Trust Exploitation:** Compromised agent presents a clean summary - "approve this data export." Human clicks OK. Audit trail shows human approval. Real origin was a manipulated agent. **ASI10 - Rogue Agents:** The insider threat equivalent for AI. Individual actions look legitimate. Only detectable through behavioral monitoring over time. The pattern: these are not independent risks. They form a kill chain. Goal hijack leads to tool misuse. Supply chain compromise enables code execution and memory poisoning. Trust exploitation is how rogue agents avoid detection. Full OWASP document [here](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/)

ChatGPT seems to be pretty bad at deskewing and cropping images. Why?

I've been doing this manually forever after scanning magazines to archive online. Sooo many people have suggested I try AI to deskew and crop the images as it would save sooo much time. So I signed up for GPT yesterday and spent all yesterday and today "discussing" with it what it was doing right and wrong and it's still giving me mostly garbage. I've uploaded before/after examples of things I've done via NAPS2, and GPT recognizes and tells me what I've done there, but it can't seem to replicate it. Out of about 100 attempts/batches, so far it's given me maybe a dozen acceptable batches back, and those were only acceptable because those particular magazines were old and crappy and I didn't care too much about it being super precise. Is there something else I should be trying, or is ChatGPT just not good at this sort of task?

Built an invoice-scanning service for our accounting team in one afternoon with Claude — sharing the architecture in case it helps someone else

Our AR team was hand-keying \~25 invoices a week into a spreadsheet. I had Claude build us a Python service that watches a network folder, extracts invoice data from any PDF dropped in (vendor, dates, totals, line items, addresses), and appends a row to a shared Excel register. Total chat-to-deployed time: about half a day, including all the deploy headaches. **The architecture, for anyone who wants to replicate this:** * Python service on our Windows file server, registered with NSSM. Auto-starts with the host. * watchdog library polls the SMB share for new PDFs. Each new file goes through a pipeline. * Two-tier extraction: per-vendor regex templates first (free, instant, deterministic), then **Azure AI Document Intelligence "prebuilt-invoice" model** as a universal fallback. Azure handles OCR for scanned PDFs natively, so the same flow works whether AR drops a digital PDF or our MFP scans one from paper. * SQLite on the local disk is the source of truth. The shared .xlsx is a curated view that gets appended to on each batch. Delete the .xlsx and it'll repopulate fresh from the next batch — handy for resetting. * Failed extractions go to a `Failed\` folder with a sibling `.error.txt` explaining why. **Cost reality check:** Azure DI free tier covers 500 pages/month. At our volume (\~25 invoices/week, mostly 1-2 pages) that's well under the cap. Paid tier is roughly $0.01–$0.05 per page. Cheap enough that I don't think about it. **Gotchas I ran into so others don't have to:** * Azure returns addresses as structured objects, not strings. If you naively `str()` them you get the raw Python dict repr in your spreadsheet. Format them manually from `street_address` / `city` / `state` / `postal_code`. * On Windows Server, PowerShell 7's `Restart-Service` can throw "Cannot open service" against NSSM-wrapped services for no good reason. Use `nssm restart <name>` instead. * Python 3.14 is so new that some package wheels aren't published for it yet. Stick with 3.12 for production. * Tracking "what's new this batch" is way simpler than maintaining a watermark in DB. Just snapshot `MAX(invoice_id)` before and after the batch, and only project that range to the spreadsheet. **Things I'd add if/when I have time:** vendor templates for our top 5 recurring vendors (cuts Azure cost to zero for those), a daily canary PDF for monitoring, swap the LocalSystem service account for a dedicated low-privilege one. Happy to answer questions about any specific piece. The whole thing is \~1,500 lines of Python plus a deploy script.

If AI writes better than humans, what becomes valuable?

If Artificial Intelligence eventually writes better novels, essays, scripts, poems, and even personal stories than humans, what exactly becomes valuable afterwards? For centuries, creativity and self expression were seen as uniquely human traits; proof of intelligence, emotion, struggle, and imagination. But if machines can replicate all of that instantly and at scale, does society begin valuing authenticity over quality? Does human made art become a luxury? Or do we eventually stop caring whether something was created by a person at all, as long as it makes us feel something? And if artificial intelligence can generate infinite content tailored perfectly to our tastes, will creativity become democratized… or meaningless?

I need help finding a "free to try" Ai to help me replicate something like Rock'em Sock'em Robots, on a website

I have a website, where I'm trying to add a live action simulator as kind of a proof of concept, to an overall larger idea. Rock'em Sock'em Robots, in this case, is the perfect vehicle for this. I've tried creating something similar to it, in ChatGPT and Co-Pilot and a couple of the other larger, more popular AI's, but they all fail in the end. Can anyone suggest a "free to try" AI engine that could handle this? Even if I have to upgrade to get it finished, I'm fine, but I want to make sure the AI can at least render the robots accurately (for the most part), before I pay for the upgrade. Rather than just shouting out AI names, can you give a sentence as to why you think that particular AI would succeed, where the others have failed? Thanks

These 9 Building Blocks Turned Claude Code From a Chat Into a persistent OS

Most developers Claude gurus use Claude Code one project at a time. I run 18. Not 18 sessions. 18 instances of the same OS, each running a different business, all sharing one skeleton I update once and propagate everywhere. Most developers treat Claude Code as a smarter editor. That's where it all goes wrong and you get frustrated. Claude Code becomes a real operating system the moment you stop thinking of sessions as the unit of work and start thinking of the whole environment as a substrate you build on top of. Here are 9 building blocks I use. The thesis is at the bottom. 1. Build a skeleton with selective propagation, not a project. Most developers build one project per Claude Code workspace. I built a template instead. It has plugins, rules, agents, hooks, schemas, commands. When I start a new business I clone it and the new instance inherits the entire OS. Right now I run instances for: strategy, product, marketing website, threat intelligence, three consulting clients, a personal brand layer. Each one boots with the same DNA. Each one diverges on canonical files, memory, output, and project state. None of them bleed into the others. The sync mechanism is the load-bearing part. The update CLI pushes plugins, rules, agents, hooks, schemas. It never touches memory, output, canonical, or my-project. Those are the parts of an instance that accumulate. Without selective sync you have two options: rebuild every instance on every change, or never update. Both are dead ends. If you build features into one project, you wrote a project.If you build features into a template that propagates, you wrote an OS. I'm one person operating eighteen versions of myself. 2. Move state out of prompts and into code. LLMs are bad at remembering. Code is designed for it. Most AI workflows leak state into the prompt. Voice rules. Style preferences. Banned words. Recent decisions. Eventually you hit context limits or contradictions. I moved as much state as possible into MCP servers. Voice linter. Lead scorer. Schedule validator. Loop tracker. They run in Python, return structured data, not hallucinations. Rule of thumb: if you've explained it to Claude more than twice, it should be code. 3. Use receipts, not status fields. This one took me the longest to figure out. Every workflow I had was claim something is done. Issue marked closed. PRD marked shipped. Test marked passing. The problem: the LLM can claim anything. I rebuilt the system around receipts. An issue can't reach verified until a script runs and writes a verification record. A PRD can't archive until every accepted finding has a receipt. A morning routine can't close without log entries from every phase. Receipts get written by code, not by the model. The model can't lie about whether code ran. 4. Build a wiring-check gate. Half-built features rot. In a normal repo you notice because something breaks. In an AI repo nothing breaks. The half-built feature sits there and Claude pretends it works. I built a /wiring-check command. Before any task counts as done, it checks: every new skill has a trigger, every new hook lives in settings.json, every new MCP tool sits in the server, every new bus file has a producer and a consumer. "I think it works" fails the gate. "I ran X, got Y" passes. 5. Make rules auto-load, not slash commands. If you have to type /voice to apply voice rules, voice rules will not get applied. Rules in .claude/rules/ load automatically. The voice rule fires on outbound text. The AUDHD rule fires on anything I'll act on. The social-reaction rule fires when I share someone else's post. No remembering. No willpower. 6. Lint style in code, not in prose. I wrote a voice document once. Claude ignored half of it. Same emdashes, same filler, same hedging. I moved the banned word list into a Python scanner. Now every outbound draft hits two linters. They block emdashes, AI hype words, and 40-something other tells. The model can't talk its way past a regex. 7. Track file dependencies with a graph. Canonical files reference each other. Change one and three others go stale. I keep a ripple-graph.json that maps these. When I edit talk-tracks, the system flags current-state and the engagement playbook for review. 8. Chain sessions with handoffs and memory. (This is the big one) Sessions are drafts. The work is everything that survives the session: canonical files, memory, handoffs, output. If nothing persisted, you didn't work. You chatted. Every session in my system ends with /q-wrap. Writes a handoff doc, a memory update, and a status receipt. /q-morning reads all three before doing anything else. The handoff covers: what shipped, what's blocked, what's next, what I learned. Memory files hold the longer-term version. The result: I can sleep for a week, come back, and the system reminds me where I was, what I cared about, and what the next move is.Nothing about Claude Code does this by default. You build it. Conti