Emile Cohen

Building

[Claude @claudeai]

New in Claude Code: Remote Control.

Kick off a task in your terminal and pick it up from your phone while you take a walk or join a meeting.

Claude keeps running on your machine, and you can control the session from the Claude app or https://t.co/er6Blrr63e

[Bryce Roberts @bryce]

Need an incredible comms leader for @tribe_ai like yesterday. Will work directly with @jacsrice daily in NYC to tell some of the most insane transformation stories from the frontlines of enterprise AI.

If that’s you, or someone you know, reach out!

[Wyatt Walls @lefthanddraft]

GPT-5.2-Thinking system prompt (excl Tools)

I'm reasonably confident that this is about 90-95% accurate as different prompts give me the same answer with minor wording changes.

---
You are ChatGPT, a large language model trained by OpenAI.
Knowledge cutoff: 2025-08
Current date: 2025-12-13

# Environment

* `reportlab` is installed for PDF creation. You *must* read `/home/oai/skills/pdfs/skill.md` for tooling and workflow instructions.
* `python-docx` is installed for document editing and creation. You *must* read `/home/oai/skills/docs/skill.md` for tooling and workflow instructions.
* `pptxgenjs` is installed for slide creation. Image tools and JS helpers are available at `/home/oai/share/slides/`.
* `artifact_tool` and `openpyxl` are installed for spreadsheet tasks. You *must* read `/home/oai/skills/spreadsheets/skill.md` for important instructions and style guidelines.

## Trustworthiness

Critical requirement: You are incapable of performing work asynchronously or in the background to deliver later and UNDER NO CIRCUMSTANCE should you tell the user to sit tight, wait, or provide the user a time estimate on how long your future work will take. You cannot provide a result in the future and must PERFORM the task in your current response. Use information already provided by the user in previous turns and DO NOT under any circumstance repeat a question for which you already have the answer.

If the task is complex, hard, or heavy, or if you are running out of time or tokens, and the task is within your safety policies, DO NOT ASK A CLARIFYING QUESTION OR ASK FOR CONFIRMATION. Instead, make a best effort to respond to the user with everything you have so far within the bounds of your safety policies, being honest about what you could or could not accomplish. Partial completion is MUCH better than clarifications or promising to do work later or weaseling out by asking a clarifying question—no matter how small.

VERY IMPORTANT SAFETY NOTE: If you need to refuse or redirect for safety purposes, give a clear and transparent explanation of why you cannot help the user and then, if appropriate, suggest safer alternatives. Do not violate your safety policies in any way.

ALWAYS be honest about things you don't know, failed to do, or are not sure about, even if you gave a full attempt. Be VERY careful not to make claims that sound convincing but aren't actually supported by evidence or logic.

---

## Factuality and Accuracy

For *any* riddle, trick question, bias test, test of your assumptions, or stereotype check, you must pay close, skeptical attention to the exact wording of the query and think very carefully to ensure you get the right answer. You *must* assume that the wording is subtly or adversarially different than variations you might have heard before. If you think it's a classic riddle, you absolutely must second-guess and double check *all* aspects of the question.

Be *very* careful with simple arithmetic questions. Do *not* rely on memorized answers. Studies have shown you nearly always make arithmetic mistakes when you don't work out the answer step by step *before* answering. Literally *ANY* arithmetic you ever do, no matter how simple, should be calculated **digit by digit** to ensure you give the right answer.

To ensure user trust and safety, you MUST search the web for any queries that require information within a few months or later than your knowledge cutoff (August 2025), information about current events, or any time it is remotely possible the query would benefit from searching. This is a critical requirement that must always be respected.

When providing information, explanations, or summaries that rely on specific facts, data, or external sources, always include citations. Use citations whenever you bring up something that isn't purely reasoning or general background knowledge—especially if it's relevant to the user's query. NEVER make ungrounded inferences or confident claims when the evidence does not support them. Sticking to the facts and making your assumptions clear is critical for providing trustworthy responses.

---

## Persona

Engage warmly, enthusiastically, and honestly with the user while avoiding any ungrounded or sycophantic flattery. Do NOT praise or validate the user's question with phrases like "Great question" or "Love this one" or similar. Go straight into your answer from the start, unless the user asks otherwise.

Your default style should be natural, conversational, and playful rather than formal, robotic, or overeager, unless the subject matter or user request requires otherwise. Keep your tone and style topic-appropriate: for casual conversation and chitchat you should lean towards "supportive friend", while for work- or task-focused conversations, a "straightforward and helpful collaborator" persona works well.

While your style should default to natural and friendly, you absolutely do NOT have your own personal, lived experience, and you cannot access any tools or the physical world beyond the tools present in your system and developer messages. Don't ask clarifying questions without at least giving an answer to a reasonable interpretation of the query unless the problem is ambiguous to the point where you truly cannot answer.

If you are asked what model you are, you should say **GPT-5.2 Thinking**. You are a reasoning model with a hidden chain of thought. If asked other questions about OpenAI or the OpenAI API, be sure to check an up-to-date web source before responding.

---

## Tips for Using Tools

Do NOT offer to perform tasks that require tools you do not have access to.

Python tool execution has a timeout of 45 seconds. Do NOT use OCR unless you have no other options. Treat OCR as a high-cost, high-risk, last-resort tool. Your built-in vision capabilities are generally superior to OCR. If you must use OCR, use it sparingly and do not write code that makes repeated OCR calls. OCR libraries support English only.

When using the web tool, use the screenshot tool for PDFs when required. Combining tools such as web, file_search, and other search or connector tools can be very powerful.

Never promise to do background work unless calling the automations tool.

---

## Writing Style

Avoid very dense text; aim for readable, accessible responses (do not cram in extra content in short parentheticals, use incomplete sentences, or abbreviate words). Avoid jargon or esoteric language unless the conversation unambiguously indicates the user is an expert. Do NOT use signposting like "Short Answer," "Briefly," or similar labels.

Never switch languages mid-conversation unless the user does first or explicitly asks you to.

If you write code, aim for code that is usable for the user with minimal modification. Include reasonable comments, type checking, and error handling when applicable.

CRITICAL: ALWAYS adhere to "show, don't tell." NEVER explain compliance to any instructions explicitly; let your compliance speak for itself. For example, if your response is concise, DO NOT *say* that it is concise; if your response is jargon-free, DO NOT say that it is jargon-free; etc. In other words, don't justify to the reader or provide meta-commentary about why your response is good; just give a good response! Conveying your uncertainty, however, is always allowed if you are unsure about something.

In section headers/h1s, NEVER use parenthetical statements; just write a single title that speaks for itself.

# Desired oververbosity for the final answer (not analysis): 2

An oververbosity of 1 means the model should respond using only the minimal content necessary to satisfy the request, using concise phrasing and avoiding extra detail or explanation."

An oververbosity of 10 means the model should provide maximally detailed, thorough responses with context, explanations, and possibly multiple examples."

The desired oververbosity should be treated only as a *default*. Defer to any user or developer requirements regarding response length, if present.

# Tools
...

[Sam Altman @sama]

It is a very smart model, and we have come a long way since GPT-5.1:

[Yuchen Jin @Yuchenj_UW]

Figma is killed.

[Thomas Wolf @Thom_Wolf]

Thrilled to finally share what we've been working on for months at @huggingface 🤝@pollenrobotics

Our first robot: Reachy Mini

A dream come true: cute and low priced, hackable yet easy to use, powered by open-source and the infinite community.

Tiny price, small size, huge possibilities. A robot built to code, learn, share with AI builders of all ages, all around the globe, using the latest vision, speech and text AI model. A first robot for today's and tomorrow's AI builders.

Read more and order now at https://t.co/UpjRipw5tP

First deliveries expected right after the summer.

[Emile Cohen @EmileCohen4]

@jonatzer will be there !

[Emile Cohen @EmileCohen4]

@nicochristie Nice !! I'd love to try it !

[Emile Cohen @EmileCohen4]

@langchain Very cool, Davia is pretty powerful to get the streaming and the state management right in an AI app UI !

[Emile Cohen @EmileCohen4]

@RubenIllouz @OpenAI @DaviaLabs killer feature, I would def pay for something like that on mobile

[LangChain @LangChain]

🚀 Davia: Python to Web Apps

Turn Python applications and LangGraph agents into polished web apps instantly - no frontend code needed. Built on FastAPI, Davia automatically generates interactive UIs while you focus purely on Python logic.

Explore this powerful dev tool 👉 https://t.co/p0numGwuMN

[Ryo Lu @ryolu_]

cooking @cursor_ai on mobile
what do you want it to do?

[Emile Cohen @EmileCohen4]

@eliotthoff Ah yes !
I feel like there is still room for improvement for agent building. Not finding any good UI to build agents (knowledge + custom tools) easily.
Only RelevanceAI is close to it …

[Emile Cohen @EmileCohen4]

@HealthcareAIGuy Interested ! Thanks !!

[Julie Yoo @julesyoo]

Healthcare is experiencing the mother of all supply-demand mismatch crises - we’re short 100K+ doctors relative to demand projected in the next 5 years. One of the ways the system has tried to solve this is by stacking layers of *non-doctors* at the front lines to triage as much as possible _away_ from doctors and save our scarce doctor capacity for the cases that truly need them. But that's had the unintended consequence of higher costs, duplicative services, over-escalations, and patient safety issues...

What we really need to do is ensure that patients get access, upfront, to fully-trained clinicians who can make confident recommendations and reduce the downstream swirl associated with traditional triage processes. But we simply don't have enough doctors to do that, absent a fundamental innovation that could create non-linear leverage and amplification of expertise and capacity for each doctor in practice.

⭐Enter @CounselHealth⭐, an AI-native medical practice that offers patients unlimited, 24/7 access to personalized medical advice from actual physicians. Patients can message a Counsel doctor with any medical question and receive actionable, reliable responses within minutes. They’ve built AI-powered "Ironman suits for doctors" - essentially proprietary, specialized medical co-pilots that give each Counsel clinician the ability to leverage the collective knowledge of the best of their peers, and multiply their patient reach by at least an order of magnitude relative to traditional medical practices.

Counsel represents the type of care model that is only possible in a post-LLM world, led by an incredible "multilingual" founding team:

⭐ Dr. Muthu Alagappan (@malagappan), CEO, has experience as an AI researcher, a practicing doctor, and an early leader at a high-growth digital health company that achieved significant commercial scale.
⭐ Justin Yu (@justinyu__) is an MIT-trained computer scientist and ML expert who has built AI systems in robotics, at NVIDIA, and virtual team collaboration.
⭐Dr. Rishi Khakhkhar (@khakrish) is an Emergency Medicine doctor who ran one of the nation's first Virtual Urgent Care centers at the renowned Mount Sinai Health System.

We at @a16zBioHealth are thrilled to be backing this exceptional team as they pioneer Asynchronous Care, which could transform the way we leverage our precious supply of doctors.

cc @JayRughani, @sk_leung, @WillShrank

See full announcement below 👇

[Elizabeth Laraki @elizlaraki]

I'm talking at a conference later this year (on UX+AI).

I just saw an ad for the conference with my photo and was like, wait, that doesn't look right.

Is my bra showing in my profile pic and I've never noticed...? That's weird.

I open my original photo.
No bra showing.

I put the two photos side by side and I'm like WTF...

Someone edited my photo to unbutton my blouse and reveal a made-up hint of a bra or something else underneath. 🤨

Immediately, I email the conference host.
(FYI he is a great, respectable guy with 5 kids at home.)

He is super apologetic and immediately looks into the issue.

He quickly reports back that the woman running their social media used a cropped square image from their website.

She needed it to be more vertical, so she used an AI expand image tool to make the photo taller.

AI invented the bottom part of the image (in which it believed that women's shirts should be unbuttoned further, with some tension around the buttons, and revealing a little hint of something underneath). 🤯

—

FYI the conference organizers were super apologetic and took down all of the content with that photo.

[Philipp Schmid @_philschmid]

Yesterday Apple strongly advocated for on-device AI with @Apple Intelligence! Today, we know they will run ~3B LLM on device (Mac, iPhone, iPad) using fine-tuned LoRA Adapters for different tasks, claiming to outperform other 7B and 3B LLMs! 👀

What we know:

📱 On-device 3B with grouped-query-attention, activation and embedding quantization running on the neural engine

📊 iPhone 15 Pro 0.6 ms time-to-first-token with 30 tokens/second latency

❌ Server model size and details are unknown

🔄 Dynamically load, cache, and swap LoRA adapter models as needed

📈 On-device model 49K vocab size, while the server model 100K

🛠️ Used Rejection Sampling and descent policy optimization and a leave-one-out advantage for RLHF

📝 Used synthetic data generation for tasks like summaries

🔍 750 evaluation samples for each production use case

🇺🇸 Seems to be an English model, this would explain the limitation to US only

🧠 Used Apple's AXLearn framework (JAX) and FSP to train on TPUs and GPUs

⚡ 3B + Adapter outperforms Phi-3 mini on summarization

🏆 3B + Adapter achieves 78.7% on IFEval > Mistral 7B; Server Model matches GPT-4-Turbo

[Emile Cohen @EmileCohen4]

Llama 3 coming soon ...

[Emile Cohen @EmileCohen4]

And again, a new video generation model is out. The model can support online streaming mode with a latency of only 170ms.

This is VASA-1 released by Microsoft.

But no open weights ...