PH热榜 | 2026-05-17

@0xaron Congrats team! Curious how memory works across long-running strategies. If an agent has been live for 90+ days, does it maintain a structured thesis/history of why it entered positions, or does it mostly rely on fresh market data and rules?

@0xaron  @abhiranjan_mehta Great question, and you've identified exactly the nuance that makes Polymarket interesting and tricky at the same time.

The agent doesn't just chase odds. Before entering a market, it evaluates signal strength, liquidity depth, and spread, thin markets with wide spreads get filtered out because the slippage kills the edge before you even open the position.

The edge vs. spot crypto is in the structure of the market itself. Polymarket resolves on real-world outcomes - so the signal layer is fundamentally different. You're not reading price action, you're reading information asymmetry. When the agent finds a market where public odds haven't caught up to what the data suggests, that's the entry.

It's a different muscle than spot trading. Took us a while to get the workflow right, but it's one of our favorite things we've built.

@byalexai Thank you for hunting us and for getting it so precisely!

You've nailed the thing we obsess over internally - research without execution is just a newsletter. The full workflow is the product.

Polymarket was a deliberate bet. Prediction markets are one of the few places where signal quality gets priced in real time. It's a natural fit for what we're building.

And yes, cost per run matters. An agent you can't afford to run daily isn't really an agent.

Excited to have the community put it through its paces today.

@jaipandya Thanks a ton!

@jaipandya ❤️❤️

@lakshminath_dondeti You’re right that some frontier model providers’ terms/policies treat finance as a sensitive or high-risk domain. But I don’t think the clean reading is “financial or crypto research / trade is prohibited.”

OpenAI’s policy calls out “tailored advice that requires a license” and “automation of high-stakes decisions” in “financial activities and credit,” not financial research broadly. Anthropic is even more explicit: finance is a “High-Risk Use Case,” including investment advice and financial eligibility/creditworthiness, but the requirement is human review and disclosure rather than a blanket ban. Google’s policy has the same general shape around high-risk finance decisions.

Our product follows all standard guidelines for such high risk use case, including disclosures, risk understandings and warnings at necessary steps.

Regarding our choice of models - It varies. We do use the top models in both open and closed weights across different places.

@suyash_kr You're thinking about it exactly the way we do.

Short answer: yes, we're going there.

Longer answer: we're actually launching our AI Quant next week - your personal quant trader that lets you build and run your own strategies, not just use ours. Swing and day trading strategies are already live, and we're seeing a 79% win rate on BTC, ETH, and SOL.

On perps specifically, Hyperliquid integration is built. Not publicly live yet, but it's coming very soon.

We didn't stop at spot because it was easier. We stopped there first because getting execution right on spot is the foundation.

Leverage on top of a shaky execution layer is just a faster way to blow up.

The foundation is solid. Perps are next.

@akshitverma5 This made our day, genuinely.

AI Quant for HyperLiquid next week is going to be worth the wait. Build your own strategies, run them autonomously, your edge your way.

Stay close. 🙏

@akhilbvs Thank you! Feels surreal to finally have it out in the world.

Would love to hear what you think once you've had a chance to try it. Your review matters!

@ishu86 This is the right question and most people are not asking it yet.

Task drift and silent failure in multi-step flows is genuinely the hardest infrastructure problem we deal with. Our approach has three layers.

First, specialist sub-agents. Instead of one general agent trying to hold context across a long horizon, each sub-agent owns a narrow, well-defined task. Smaller scope means less surface area for drift.

Second, ordered task execution with checkpoints. Fere breaks a strategy into a sequence of atomic tasks. Each step validates its own output before passing to the next. If something looks off, it does not silently proceed.

Third, live market feedback as a correction signal. Because our agents operate with real wallets in real markets, they get continuous ground truth. Reinforcement learning against live outcomes means errors surface fast and the agent self-corrects over time rather than compounding mistakes.

We will not claim we have fully solved this. Nobody has. But 10 million live executions across 90 plus day strategies gives us a real feedback loop that most agent products running on synthetic benchmarks simply do not have.

Would love to compare notes on what you are seeing on your end.

@imrulkaayes Love this question. For spot DEX swaps, we aggregate the aggregators — pulling from nearly all the top ones by volume across every chain we support. The optimizer balances execution reliability and best price in real time. So yes, exactly like a trader frantically tabbing between aggregator UIs during a volatile candle, except Fere does it deterministically in the background. Manual execution leaves alpha on the table; we close that gap.

@vishalmehta8340 Not exactly in the way you described but we thought about this a lot.

A confirmation-before-every-trade mode sounds great on paper, but the token cost per decision makes it genuinely impractical at scale. We'd either have to charge a lot more or the agent would be too slow to be useful.

What we do instead: you can backtest and simulate your strategy first. Run it, see how it performs historically, get confident in it, and then flip it to live execution. Same outcome, without the overhead.

And for users who don't want to build their own, our pre-built strategies have been tested and have been delivering consistently. You're not flying blind either way.

@rahul_singh_bhadoriya Great question, and the one we take seriously. A few layers to how we handle it:

First, the agent doesn't just execute blindly. Risk parameters: position sizing, stop logic, exposure limits - are built into the workflow before a trade ever fires.
Second, flash crashes are a speed problem. Our execution layer is designed to react faster than a human watching a chart, but we also have circuit-breaker style controls so it doesn't double down into a free fall.

Black swans are harder - no system predicts the truly unprecedented. What we focus on is limiting downside when the model's confidence is low and the market is behaving anomalously. The agent knows when to step back.

Happy to go deeper on any specific scenario.

@conduit_design Sycophancy in LLMs is real, and a trading agent that agrees with everything you say is a liability, not a feature.

The Fere Agent Harness is our answer. The agent builds its own thesis first — grounded in research, live market data, and prior context — before it ever responds to your input. When your view contradicts the evidence, it doesn't fold. It surfaces the conflict, attempts to resolve it, and shows you the reasoning. Think verifier, not yes-man.

Final call is always yours. But by the time it reaches you, the agent has actually done the work to disagree.

@kumar_ritesh21 Not exactly, and here's the honest reason why.

A confirm-before-execute mode burns a huge amount of tokens per trade decision. It's technically possible but it makes the product expensive and slow for most users. Didn't feel like the right trade-off.

Instead we give you backtesting and simulation - run your strategy, validate it, get comfortable with how it behaves. Once you're confident, you go live.

Our pre-built strategies have already been through that process. Tested, refined, consistently delivering. So you're covered either way.

@flowcast Yeah, this is the question we obsessed over the longest before launch.

The honest answer: neither Fere nor the AI ever has access to a private key. We're built on Coinbase's CDP Server Wallets, which means every wallet — user or agent — lives inside a secure enclave (AWS Nitro). Keys are generated and used for signing inside that enclave and never leave it. Not exposed to Coinbase, not exposed to us, not exposed to the model. The agent simply makes scoped API calls; the enclave does the signing.

The other half of the answer is policy enforcement. Each wallet has programmable guardrails baked in at the signing layer — what contracts can be called, spending caps, address filters, security checks. So the agent isn't free to do "anything"; it operates within rails set, and the wallet itself will refuse to sign anything outside them.

That's how we square the circle: you get a seamless, one-click experience across chains, and the security model doesn't depend on trusting the AI to behave — it's enforced cryptographically.

@asti_pili You don't need to be a quant to use Fere. That's kind of the point.

The agent handles the research, the signal reading, and the execution. Your job is telling it what you want, risk level, assets, strategy style.

If you've ever bought crypto before, you have enough context to start. The rest you pick up as you go.

Where it gets more powerful is when you layer in your own preferences over time. But day one? Just dive in.

@sinchana_v Sharpest question of the day and one we've thought about a lot.

A few things work in our favour here.

First, Polymarket markets are outcome-based not price-based. The alpha isn't in being faster than other traders, it's in reading information that hasn't been priced into the odds yet. That's a fundamentally different dynamic from spot crypto where crowded signals kill returns quickly.

Second, we don't route all agents into the same market at the same time. Liquidity and spread are first-class filters. If a market is too thin to absorb position flow without moving the odds against us, the agent doesn't enter. We'd rather miss the trade than cannibilise it.

Third, and honestly the more important answer: as the user base scales, signal diversity scales too. Different users, different risk appetites, different strategy configs. Not a monolithic flow.

The crowding problem is real in quant. We're just not as exposed to it as a pure price-signal product would be.

@hunter_powabase Right now, you don't. The LLM Costs are included in the credits charged for each conversation. This means our credits per chat is dynamic. Simple stuff consumes fewer credits than complex ones.

We are considering a possibility of letting users choose their own models in a future release.

Pro Tip: You can also use our MCP or CLI to make it work with your own agent (Claude Code, Hermes, OpenClaw etc). https://docs.fereai.xyz/mcp

@hirogure Real-time, not pre-set. Our routing layer reads market conditions live and adapts as it goes. If execution doesn't clear during a volatile move, the system keeps adjusting — trying different routes and recalibrating — until the trade goes through. Pre-set slippage values tend to break exactly when you need them to work, so we built around that.

@summer_dev 感谢您的美言。我们收到的反响简直令人难以置信。

@thamibenjelloun Yes, and this was important for us to get right.

We have backtesting and simulation built in. You can run your strategy against historical data, see how it would have performed, and get comfortable with it before a single real dollar moves.

Once you're confident, you flip it live. No pressure to go autonomous on day one.

Our pre-built strategies have already been through this process extensively so if you'd rather start with those while you find your footing, that's a completely valid path too.

@thamibenjelloun  Yes — simulation and backtesting are built in, not an afterthought. Run your strategy against historical data, watch how it would have played out, adjust, and only go live when you're convinced. No real capital at risk until you've seen the evidence. It's the same philosophy behind our pre-built strategies: every one of them has a visible track record, so you can evaluate before you commit.

@zijian if a user doesn’t like one scene, can they regenerate only that scene/keyframe while keeping the rest of the story consistent?

@bsy0221  @chrismessina Now Marvel cosmo should invite Dr Arthur to their squad !

@itsluo 

Wow, thanks for the shoutout!

Coming from you, this means a ton to our engineering team.

We're super proud of how HiDream-O1-Image performs as the backbone. Beyond the model itself, we actually built an entire 'swarm of AI directors' workflow around it this time. It acts like a real creative studio to plan the scenes and map the visuals before rendering, which is how we pushed the boundary to a stable 3-minute output.

Appreciate you stopping by!

@artstavenka1 

Love that you noticed the storyboard logic!

Let me share a quick insider update on how our architecture scales up from short to long-form storytelling:

For our 15-second mode, yes, it relies heavily on that 9-frame storyboard blueprint to lock coherence. But for our full 1-minute narrative mode, we actually upgraded to a much more advanced, film-industry-style pipeline!

Instead of boxing the AI into a rigid 9-frame grid for long stories, our unified planning layer takes your story script and pre-generates a cohesive set of 'Global Production Assets' - including character sheets, environmental concepts, and key props, etc.

Our core agent framework then uses these unified reference assets as absolute anchor points to guide the continuous rendering of the full 1-minute video straight from the script. This gives our workflow much more creative freedom for dynamic camera tracking while ensuring zero character drift. It’s total control, just scaled up for true cinematic production! 🎬

@lakshminath_dondeti 

Great question!

Traditionally, you have to describe every frame, camera angle, and style parameter yourself. With Vivago, you skip all of that.

You just type a simple story idea in natural language. Then, our built-in 'swarm of AI directors' takes over. They automatically brainstorm the characters, break down the scenes, and write a cohesive creative logic for you.

Before the heavy rendering begins, the tool presents a 9-frame storyboard preview so you can see exactly what the creative direction looks like. It gives you 100% confidence in the output without doing the heavy lifting of prompt engineering!

@christian_knaut 

You absolutely understand the user psychological barrier! The '20 failed attempts before quitting' is exactly the pain point that kept us up at night. 😭

To answer your question:

we built the workflow to adapt completely to your style of creative direction. You can be as hands-off or as hands-on as you want, using pure natural language:

• The 'Hands-Off' Mode: You can literally just give it a loose topic or a vibe (e.g., "Make a comedy sketch about a cat trying to code"). Our agent swarm will jump in, brainstorm the script, set up the scene environments, and handle the heavy lifting for you from scratch.

• The 'Director' Mode: If you have a specific vision, you can feed it a full script, upload reference images for your characters, scenes, or props.

And you actually anticipated our roadmap perfectly with the mention of reference videos!

While we currently focus on text and image-based assets to anchor the continuous narrative, supporting reference videos for advanced motion and camera guidance is officially coming next. Stay tuned for that! 😉

The best part? You don't have to get it perfect on the first try. Because our system supports continuous natural language dialogue, you can just chat with the AI director to iterate, tweak, and refine any specific stage of the production pipeline on the fly. Go grab the 1-month free Plus code on our page and see how it feels. You control the narrative, we handle the execution! 🎬

@owen_shaw2 

Thank you for the congrats!

You just hit on one of the most critical engineering challenges in commercial AI video. 🎯

To give you the exact technical breakdown of how Vivago tackles this: we built a specialized pipeline that combines Vision-Language Models (VLM) with an internal 'Asset ID' tagging system.

Instead of just blindly projecting your brand logo or fonts across frames, our planning layer uses VLM to deeply comprehend the graphic identity, geometry, and semantic meaning of your upload. It then assigns a unique, immutable Asset ID to it.

Throughout the long-context generation of a 1-minute video, our core agents track this Asset ID continuously, instructing the backend to precisely anchor and lock the IP’s visual consistency without distortion. It’s a deterministic, studio-level approach to brand safety in AI! ⚡

@imogen_wallace 

Thank you so much for the congrats! You just hit on one of my favorite features. ❤️

YES, you absolutely can use your own character designs! In fact, I just used photos of myself and my cats to generate my own story earlier today.

While our 'swarm' is brilliant at brainstorming characters from text, we designed Vivago specifically for real creators who already have fixed IPs. For our long-form mode, you can upload your own character images, reference sheets, or environmental designs right into the pipeline.

Our planning layer locks those visual identities first, and then the agent swarm directs the rest of the storytelling around YOUR assets. This ensures your character stays consistent across every single frame. Go grab the 1-month free Pro code on our page and bring your own characters to life! 🎬⚡

It won't take you too long, so you are more than welcome to post your video onto YouTube & share the link here with the community!

@monir_ 

Haha, I totally respect the skepticism! The AI space is flooded with hyper-curated clips, so it's 100% fair to ask 'what's the catch?'

To be honest, we are so confident in our backbone that I'm not even going to waste your time trying to convince you with another official demo. Instead, I challenge you to go break it yourself. 😉

Get 1-month free membership with code "1MPRO4PH", upload a photo of yourself, and describe your story.

You can retry the same description & images we used for this demo video as well:

Description: A kitten travels through a highly futuristic digital frontier, searching for Arthur and inviting him to become the legendary Product Hunter for the global launch of the vivago Agent website.

Chris (our No. 1 hunter) profile photo can be found here: https://messina.as.me/schedule/9ce1ce24/appointment/60975938/calendar/9937930

Product Hunt Kitty image we used is this one:

https://elev.io/blog/featured-on-product-hunt-what-we-did-right-and-what-wed-change

The only real 'catch' right now is that a full 1-minute narrative takes about 40-50 mins because the agent swarm is doing heavy frame-by-frame coherence plotting. But the output is raw and real.

Let us know if we passed your BS test!🚀

@1mirul 

Thanks for the love! Glad you like the smoothness. 🙌

Just a quick clarification to keep expectations aligned: Vivago currently renders native 1080P video rather than 4K.

Regarding the rendering speed:

a standard 15-second video takes about 10 minutes, while a more complex 1-minute narrative takes around 40 to 50 minutes.

Because our 'AI Directors' are doing heavy, frame-by-frame orchestration and visual continuity alignment behind the scenes, it takes a bit of time to bake.

But we promise the coherent quality is absolutely worth the wait! 🎬

@abod_rehman 

Appreciate the sharp question!

Vivago is built from the ground up as an agent-native product.

We’ve architected a proprietary three-tier agent system consisting of 'Tools + Skills + Workflows,' all orchestrated by a single, unified planning layer that schedules all core capabilities dynamically.

The magic is that our agent framework is natively integrated right into the core architecture of the HiDream-O1 model itself.

So rather than a chaotic struggle between model power and prompt layers, it’s a seamless synergy where the model naturally understands and executes the structured agent logic.

@boyuan_deng1 

You nailed the exact core of our product vision!

Yes, iteration is absolutely at the heart of our AI Director workflow.

For our standard 15-second creation mode, we fully support up to 10 rounds of back-and-forth dialogue to freely tweak and modify the generated content.

And it gets even better. For our longer 3-minute storytelling mode (currently in Beta), we’ve unlocked a completely unrestricted, natural language dialogue system. This allows you to interactively guide, iterate, and refine any specific stage or frame of the production process just by talking to it.

No starting from scratch required!

@cheng_ju1 

Thank you for recognizing the journey!

Two years is a long time in AI, and we’ve poured countless hours into reimagining what the future of video production should look like.

We are thrilled that you love the new structured workflow. Our design goal was to remove the randomness of text prompts and give creators a predictable, step-by-step framework through our AI directors.

Having a long-time user validate this design evolution is the ultimate reward for our product team.

Thanks a lot for the continuous support over the years! So glad to have you back.

@matthias_bettin 

Hey there! As the Product Manager for Vivago, reading your comment put a huge smile on my face. You’ve basically predicted our next major feature roadmap! 🔮

You are 100% right! The traditional 'one-click blind box' video generation leaves too much to chance. That’s why we built Step 3 (Preview Keyframes) into our current workflow.

But we want to take it even further. We are currently dogfooding an upcoming feature internally called 'Plan Mode'.

In this mode, you will get a dedicated creative sandbox before the video generation even begins. You can freely prompt, generate a bunch of diverse images, pick your absolute favorites, and use them to lock in the visual style. The Video Agent then takes those exact selected images and animates them into a continuous up to 3 minutes story.

It’s going to be exactly that 'Grok Imagine but better' experience with full creative control. We are pushing hard to ship this in our next big release.

Can't wait to hear what you think when it's live!

@jim_jeffers 

You are reading our minds! What you just described is exactly what our engineering and product teams are actively dogfooding behind the scenes.

We are currently internal testing an independent memory system designed precisely for this. Right now, it operates similarly to a dynamic design.md config file - a centralized source of truth that absorbs, remembers, and syncs your latest brand sense and visual philosophy across projects.

Looking ahead on our roadmap, we will be opening up this memory system to give teams direct editing power. You'll be able to micro-adjust the AI's internal 'Brand Bible' on the fly to ensure absolute, unbreakable brand consistency.

You've perfectly articulated the core problem we want to solve for enterprise teams. Stay tuned, because this 'creative memory' is coming to life very soon! 🎉

@yashika_vahi 

Thank you so much! As a social media strategist, you’ve pointed out the absolute holy grail of commercial video production. 🎬🎯

To answer your question: Vivago handles brand guidelines and color palettes through our Structured Creative Process. During the initial 'Share Assets' and 'Character Invention' phases, you aren't just giving the agent a prompt, you are establishing a Visual Anchor.

Our Swarm of AI Directors locks in these brand assets, character sheets, and specific hex codes before the narrative generation begins. The agents then collaborate to map out the storyboards based on those exact parameters, ensuring that the color palette and character features stay completely on-brand across all scenes.

Think of it as passing a strict brand book to a human production team - our agents review and enforce it at every keyframe. We'd love for you to stress-test this with your own brand guidelines and let us know how it holds up. 😁

@thamibenjelloun 

Thank you so much! Thrilled to hear that you noticed the quality upgrade.

We really pushed our engineering limits to deliver consistent 1080P narrative videos this time. Can't wait for you to experience our swarm of AI directors.

Let us know what you think!

@madalina_barbu Thanks a lot! 🔥 Appreciate the support!

@bhawna_rathee Great question! We put a lot of craft into the generation pipeline, things like multiple fact-check layers, web search with sourced citations, and structured orchestration rather than a single raw prompt-to-audio pass. We're actually about to publish a benchmark report on quality and accuracy, so stay tuned for that!

@david_richard35 Exactly this. Audio fits into life without asking you to carve out new time. Glad it resonates!

@samir_asadov your framing is exactly right and your topic sounds like a perfect audio native slice. Would love to hear what you think after you try it!

@lily_liu8 Yes! Custom source material is on our roadmap, your PDFs, articles, URLs. For now you can paste content directly into your prompt as a workaround. Stay tuned!

@mia_y Love hearing this! RAG architecture is exactly the kind of dense technical topic that clicks in audio format. Glad the setup was smooth!

@thamibenjelloun Exactly! Generation without a destination is half the job. The "where do I actually listen to this" problem is what we kept hearing, so Spotify felt like the obvious answer. Hope you give it a try!

@hanzhizhang0405 this made our day!

@andrewb23 thanks Andrew! And Walter Isaacson sounds like a perfect test case, as those books are dense and long, exactly where an audio digest shines. Would love to hear what you think after you try it!

Great catch on the API limitation and that was true until recently! Spotify actually just launched "Save to Spotify" (May 7), an official CLI built exactly for this use case. It now lets users push custom audio directly into their Spotify library as "Personal Podcasts," available across all their devices.

@anusuya_bhuyan haha we'll leave the breakup albums to someone else, as we're all about learning, briefings, and knowledge content. But hey, wild times indeed!

@gia_xu Custom source material is on our roadmap, your PDFs, articles, or Youtube URLs. For now you can paste content directly into your prompt as a workaround. Stay tuned!

@kawshar really excited to check. many congratulations.

@kawshar how does the canvas map back to clean WordPress output? Does Kirki generate standard HTML/CSS with Grid/Flexbox, or is there a custom rendering layer?

For designers, freedom is amazing — but performance and maintainability matter a lot too.

@ehsan_riyadh Thank you 🙏

@salmansaafi thank you 🙏

@jagadeesh37 Congratulations on your first initiative on this kind of creative work

To

Mr. Jagadeesh

creator of "CODE BREAK"

@jagadeesh37 Curious how you’re thinking about the roadmap once CodeBreak becomes universal. Would it monitor Cursor, Codex, Gemini CLI, etc. separately, or become one unified “AI work status companion” across all agent tools?

@deepak_pathania How are you generating or selecting the answer choices? Are the wrong options random legal moves, common mistakes, or intentionally tempting tactical alternatives?

@chawlaaditya8 Thanks for trying it out, see you on the leaderboard!

@hrishikesh_sharma Thanks, Hrishikesh.

Goal was to make chess puzzles feel more like a game, so it’s great to hear that came through in your experience!

@ruchi_verma4 Thanks for trying it out!

And yes, once the platform has enough players, I’ll add a matchmaking layer, with a rating system to match similarly experienced players.

Till then, please continue to enjoy playing with your friends!

@xdoneonx One feature I’d personally find useful: a “renewal week” view that shows what’s coming up in the next 7/14/30 days, so I can decide what to cancel before the charge hits.

Are you planning more planning-style views, or keeping the app intentionally minimal?

@hboon Thanks for checking us out! Great questions.

Here is how the architecture works under the hood:

Data Persistence: Everything is stored locally on your device using MMKV, which gives us blazing-fast read/write speeds compared to traditional AsyncStorage or CoreData.

Syncing: To keep things seamless without forcing you to make an account, we use iCloud sync. Your data stays yours, safely synced across your own Apple devices.

Reminders: Spot on! Reminders are triggered entirely via local notifications.

Our goal was to build something fast, lightweight, and 100% privacy-first—no tedious sign-ups or third-party servers tracking your data. Let me know if you have any other questions!

@quang_nguyen37 Let's go Quang!