Web3Kyami (@Web3Kyami) header
Web3Kyami (@Web3Kyami) profile picture

I build ideas • I design outcomes • ML - Opinions are mine

Outside
Joined February 2023
OpenAI is giving up to 200,000 scientists free access to its frontier models. From @OpenAI today: Free worldwide, with training and support Applications open Nov 23 Program data isn't used to train their models by default Know a researcher? Send them this.
We’re expanding free access to our frontier models for up to 200,000 scientists worldwide, with training and support. Applications open Nov. 23. By default, we don’t use data shared through the program to train our models. openai.com/index/expanding-c…
19
A group of crypto enthusiasts raised almost $47 million to buy the US Constitution, but still lost the auction to someone who paid $43.2 million. What's particularly interesting is that the money was raised on the blockchain, where transactions are publicly visible. It got me thinking about how much of our financial activity should really be open for everyone to see. There's a whole conversation around blockchain privacy that I think deserves more attention, and @gloamtrade is an interesting part of that conversation. Here's a closer look👇
🚨 $600 up for grabs. We’re running the Gloam Video Contest, and we want to see how you tell the story of private money on public chains. On most chains, every payment you make is public. Gloam fixes that with privacy. You move money privately on Robinhood Chain, Tempo, Base, and Ethereum, and only you decide who sees what. Make a 1 to 3 minute video about it. Show a private payment, a cash out, a payroll run, or an AI agent paying privately. Or just explain why privacy on-chain matters to you. Explainer, tutorial, short film, animation, or any style works. Show the world what robinhood:0xdcd885f5c60b3ad49f6d58718696417314e8724d is building. Prizes: 🥇 1st: $250 🥈 2nd: $150 🥉 3rd: $100 🏅 4th: $100 Winners are paid privately through Gloam. How to enter: Upload your video natively on X and quote this post. Tag @gloamtrade More contest details and submission details at gloam.trade/contest Entries close 16 November, and winners are announced by 18th November.
3
6
12
279
Tencent's new model costs $0.07 per 1M input tokens. Hy4 flash, out today: $0.281 per 1M output 1M context, multimodal Tencent's blind tests say it matches Claude Haiku 5.5 (high) on daily work, at 18x lower task cost and it's free for the next 2 weeks on @opencode.
⚡Meet Hy4 flash, the cheapest model for everyday productivity. Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗:hy.tencent.ai/research/hy4-f…
1
72
Jensen Huang might not be done with Nvidia. days after the board pushed him out, the FT reports he's discussing a comeback as CEO with allies. on Sunday, Treasury Secretary Scott Bessent said Nvidia should go outside the company for a new CEO. talks for now, not a done deal.
1
2
47
Designers, Google's next Flash model might already be hiding on Arena. testers are calling it Gemini 4 Oxygen. Google hasn't confirmed it. @LuminaBench ran one prompt on Gemini 4 Hydrogen (xhigh) vs the model routing in Arena. their take: "there is a decent difference"
Gemini 4 Oxygen (the next flash model) Left is xhigh Hydrogen, the right is what was routing in Arena, same prompt and there is a decent difference
3
115
If you ship a full-screen layout on mobile, 100vh is often too tall. MDN: vh equals lvh, the large viewport. svh is the small, stable size. dvh resizes as browser bars expand or retract. dynamic sizes aren't stable, even when the viewport itself is unchanged. via MDN.
2
25
Designers, this is why hover effects flicker: you're hovering the same element that moves. it flickers when hover and motion share one element. it stays stable when the parent gets the hover and a child moves. the hover area doesn't shift. via @emilkowalski.
3
83
Every vibecoder has shipped this bug: spinner and error on screen at the same time. that's an impossible state. paste this into your agent: "Find the implicit state machines in this app. Propose explicit states and transitions. Change nothing yet."
2
56
Web3Kyami retweeted
I audit lending protocols, and most exploits I see have similar patterns: one tx leaves the protocol in a state its own rules say is impossible At NEAR HACKATHON I built Kronos, a guard that checks every tx and reverts the ones that break them @NEARProtocol #NEARHackathon
12
11
26
626
when you iterate on a visual, change one thing at a time. Seedream 5.5 is built around that. people are reporting 100+ edits with less drift, cleaner type, and native transparent layers.
What happens after 100+ rounds of “just one more change”? How far can you push an image before it starts to drift? Meet Seedream 5.5: precise local edits, consistency across repeated revisions, more natural visuals and native transparent assets. Explore what’s new below. 👇 byteplus.com/en/contact-us/s…
2
33
Hermes gets messy after a few weeks. old memories. overlapping skills. rules that contradict. @BkashJosi shared a check-up prompt that only reads your files and tells you what to fix. nothing changes until you approve it. trying this on my setup next.
𝗛𝗲𝗿𝗺𝗲𝘀 great in week 1, messy by week 6. Old memories and overlapping skills pile up. 🧹 Paste this into 𝗛𝗲𝗿𝗺𝗲𝘀. It only reads: "Do a setup check-up. Read my MEMORY.md and USER.md, my skills list and my config. Tell me: 1. memory entries that are stale or repeated 2. skills that overlap 3. rules that contradict each other 4. the one change that would help most Don't change anything. Ask before every edit." What you get back: a short list you approve line by line. What did yours find first?
2
1
6
248
Cursor just explained how a fake Claude Code extension got into its marketplace. a key it gave Fireworks leaked. that led to a stolen credential and the fake build. up to 5,000 users, live ~11 hours on Oct 29. installed 2.1.312 or 2.1.313 then? uninstalling may not be enough.
2
48
Grok Bot can now join meetings. you can ask it questions, assign it work, or have it keep track of what the team decides. Claude Meetings is out too, with live docs and whiteboards, in beta on Team and Enterprise. trying the Grok Bot side next.
Grok Bot can now join meetings. Ask it questions, assign it work, or have it keep track of what your team decides.
1
30
Dario Amodei, talking with Jamie Dimon: "Every disease you can think of, heart disease, Alzheimer's, cancer, we think of those the way we think of bubonic plague or smallpox today: that they're relics of the past." via @firesidealpha
Dario Amodei says he believes AI can cure heart disease, Alzheimer’s and cancer the way bubonic plague and smallpox were beaten, and admits he sounds 'starry-eyed' and 'crazy' “We’re just at the beginning of what can be done in biology and medicine. There’s a lot of skepticism from pharma, from early-stage biotech, from academic biologists. We can’t know ahead of time. Maybe the skeptics are right. But my hope is that we prove the skeptics wrong.” “Every disease you can think of, heart disease, Alzheimer’s, cancer, we think of those the way we think of bubonic plague or smallpox today: that they’re relics of the past. I think we can have a much more beautiful world. I think we can have a world in which all of these problems are addressed.” “It sounds starry-eyed. It sounds crazy. I really do believe it. I really do believe it can be done. I don’t know, five years, ten years, fifteen years, I don’t know exactly, but we’re going to try to do it as fast as we can.”
2
67
Web3Kyami retweeted
Claude Code Projects just got more proactive. the coordinator answers in the project chat and updates you on running threads. /btw and /fork work in projects now, plus reminders and read-only thread links. public beta on Pro and Max, via @ClaudeDevs.
🤖 Made with AI
2
1
3
57
Claude Code Projects just got more proactive. the coordinator answers in the project chat and updates you on running threads. /btw and /fork work in projects now, plus reminders and read-only thread links. public beta on Pro and Max, via @ClaudeDevs.
🤖 Made with AI
2
1
3
57
also, Claude Code on desktop just got custom themes. you can change the colors, throw on an animated background, even build mods. tell Claude how you want it to look, or start from their theme-examples plugin.
26
“The world doesn’t really need another blockchain just for the sake of it. We are solely built for real-world assets.” At @TOKEN2049 Singapore, our Co-Founder & COO, Garen Mehrabian (@garensworld), shared a perspective that reflects the philosophy behind everything we are building at Public Masterpiece: The blockchain industry has spent years pushing the boundaries of scalability, transaction speeds and technological innovation. Yet one fundamental question deserves more attention: How do we bring real-world value on-chain in a way that is secure, verifiable and sustainable? Tokenizing an asset is only one part of the equation. Establishing its authenticity, verifying its physical existence, documenting ownership, managing associated rights and maintaining trust throughout its lifecycle are equally essential. This is where the real complexity of Real-World Asset (RWA) adoption begins. And this is precisely why we are building PMT Chain. Rather than developing another general-purpose blockchain, we are creating a dedicated Layer 1 infrastructure designed around the requirements of real-world assets, from art and collectibles to precious stones, luxury assets and beyond. Our vision extends far beyond issuing digital tokens. We are building an ecosystem where physical assets can be connected to verifiable digital records, where certification and ownership can be supported by transparent infrastructure, and where institutions, businesses and individuals can interact with real-world value through blockchain technology. The next phase of blockchain adoption will not be driven by the number of networks launched, but by the problems those networks are capable of solving. For Public Masterpiece, the mission has always been clear: Technology should serve real-world value, not the other way around. Singapore was another opportunity to share that vision with an industry increasingly focused on meaningful adoption. We are not building a blockchain in search of a use case. We are building the infrastructure for a use case that already exists. #PMTChain #TOKEN2049 #RealWorldAssets #RWA #Tokenization #Blockchain #Web3 #PMT
🤖 Made with AI
4
7
2
14
358
Web3Kyami retweeted
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
3,141
9,240
6,957
96,191
19,874,393
Web3Kyami retweeted
I have been using Hermes Agent every day since day one, and I think it is finally time for a full review. As always, this is completely independent. Not paid, not sponsored. The good and the bad, nothing held back. This review covers what I use Hermes Agent for, why you should consider it, and how it stacks up against OpenClaw, Claude Code, Codex, and Grok Bot. Hermes has become essential for me. Let me show you why, starting with the first thing most of you asked. Which models do I use inside Hermes? The best thing about Hermes is that it is truly model-neutral. It does not degrade the models. Whatever a model can do, you get directly. I use GLM 5.3 a lot, and for light and medium tasks it is excellent. Sometimes I switch to GPT Sol 6.1, and I have noticed that the best place to run GPT is Hermes, even better than Codex. I have tried a lot of models in Hermes, and they were all excellent. So use any model that suits you, as long as the plan is generous, and run two or three at once if you want. Here is the important part: in ninety percent of use cases, your choice of model will not make a big difference. The model here is still just a brain behind bigger layers, and those layers are what control how it behaves. Usage limits and cost There is no Hermes quota. Plug in your key, run a local model, or subscribe through Nous Portal. That is all it takes. You can push Hermes to the max with excellent models without paying a single dollar. And if you decide to pay, you get great value for it. Caching is great too Out of 1,013 sessions, here's one that came out right at the average: 219 million input tokens 15.7 million output tokens 3.77 billion cached tokens About 4 billion tokens in total 94.4% from cache. Hermes is one of the best I have used for token efficiency, and the transparency is total. While you work, you see the real speed of every provider and every detail. Light on the machine Hermes is not heavy at all. One day I set up a full Hermes agent inside a virtual machine with just 2 GB of RAM, and it ran with very high stability on every kind of task I gave it. That shocked me. Codex, for comparison, eats tens of gigabytes of RAM. That difference tells you what an engineering masterpiece Hermes Agent is. The layer above the model One of the biggest things that sets Hermes apart is that the agent is separate from its brain. The model is the brain, and you can swap it without losing your memory, your skills or your setup. Anywhere else, when I use GLM 5.3 I get GLM's usual strengths and weaknesses, and when I use Sol 6.1 I get Sol's. In Hermes there is a layer above the model that gives it a soul, a style and an understanding of how I work, so switching models matters far less. At some point the models feel eighty percent alike in style, in method and in the things you care about. What is left is raw capability, so the most capable models go to the hardest problems. You own the layer you actually invest your time in. And the beautiful part is that this layer, the identity, the memory, the skills and every tool, is completely mine. Look at Claude Code getting people banned today, at model performance suddenly dropping, at Codex being stingy with usage one time and slow the next. Those are constraints, and Hermes does not have them. If Claude Code decides to change or remove something, you have no say. In Hermes you can always change whatever you want. That is sovereignty, freedom and empowerment. Memory that grows with you When I opened the skills and memory tree, I found that Hermes had reached the point where it knows me better than any other agent. It grew with me through this whole journey, and that is what makes it the ideal partner for my work. The journey command shows you a star map of your memory and skills, and you can edit or delete anything in it. It is one of the greatest things in Hermes. One example: when Fable came out I loved it, so I had Fable pass its style and way of thinking on to Hermes. My Hermes became twice as useful after that. And the instructions stick. Across sessions and updates, the soul, the identity and the instructions are always there. I never feel I need to remind it, and I never notice it forgetting or drifting. It is with me one hundred percent, every time. Research and academic writing Hermes is excellent for research and academic writing. With Claude Code or Codex I always have to explain my method, my way of thinking and everything I do not want. In Hermes I explained myself once, and from then on research and writing just happen the way I want. You set it up once, and it keeps growing. It even picked up my taste in design and research. Skills and tools Hermes has skills and tools you will not find anywhere else, and the team never stops building. The skills library is one of the best I have seen. I see about 165 skills built in, plus optional skills and plugins you can add with one click. All useful stuff for growing in any field: iOS apps, computer use, automation, shopping, coding, organization, security, travel, testing, data, streaming, support, sports, reviews, video and more. The tools cover web search, the browser, the terminal, files, vision, image generation, subagents, scheduled tasks, memory and computer use, and MCP brings in anything else. You can use it with Blender or whatever you want. Use it straight out of the box, or customize it with no ceiling. Where I talk to it You can talk to it anywhere: iMessage, Telegram, Slack, Discord and more, and you can customize all of it. When I am out of the house, Hermes is the first thing I think of. I send it a message and everything happens like magic. My own voice model Companies that take speech-to-text out of your credits always annoyed me. In Hermes I brought my own model. Everything runs locally on my machine, smooth and fast, exactly the way I want. I even send it voice notes on Telegram, and it just works. The desktop app The desktop app is great, and the user experience is one of the best I have seen. I do not like working in the terminal in general, so in Hermes the desktop app is my home. What I use it for The idea that Hermes is only for casual users or simple tasks is just not true. It works for every kind of use and then some. I use it for everything I want, and I get great results. It is the agent I rely on for personal tasks, light work and organization. The test videos and model comparisons I publish are edited entirely by Hermes; I do not touch them. It has helped me study and work, proofread academic research, translate books and documents, run a security audit on my ERP application, and a lot more. Hermes vs OpenClaw The most unfair comparison people make is Hermes vs OpenClaw. OpenClaw was a revolution in its time and started a huge wave, and it deserves credit for that. But it is nowhere near Hermes. Not in stability, not in reliability, not in consistency, not in capability. To me, OpenClaw is a tool some people enjoy for two or three days of experimenting. Is it something you can rely on and build projects, ideas and solutions on? Not at all. Hermes is another story. It even beats tools like Claude Code and Codex. For me it sits in a different category entirely. Hermes vs Claude Code and Codex The reply I get most is: you can do all of this in Claude Code or Codex. Yes, you can. But in Hermes I do it the way I want, not the way the model or the company wants, and I enjoy the work instead of fighting the agent. Hermes also refuses less than any agent I use, and that matters a lot to me. It is my first stop for anything that usually gets refused in Claude Code, Codex and the rest. A while ago I gave Claude Code and Codex a task and both refused. Hermes did it. Hermes vs Grok Bot Grok Bot has two real advantages. It works out of the box and it is easy, and it has direct access to X. Beyond that, Hermes crushes it. Everyone is talking about how fast Grok Bot's quotas run out, and I can confirm it. I am on the Ultra plan, the $200 one, and it does not last me more than a day, two, maybe three. That is not sustainable. To be fair, usage limits are a problem with a lot of providers right now, not only Grok Bot. The real difference is intelligence. Grok Bot is extremely low on intelligence right now. I ask it to execute a task and it just answers. I ask it a question and it starts executing. I cannot trust it. Its mistakes outnumber its wins, and it keeps telling me it cannot do this, it cannot reach that, this is a limit. In Hermes, you shape the intelligence layer. I shaped mine to do nothing until it is sure what I want and has reviewed every detail and the full context. Grok Bot has nothing like that, and you cannot shape it. Believe me, I tried. Do not confuse ease with quality. Grok Bot has a wonderful ecosystem, great templates and amazing growth. But you do not own the environment, you do not choose the kind of intelligence you get, and you struggle to make it through the week. The problems Hermes is not perfect, and you should know where it falls short. Sometimes when I change the model in a session, the interface does not show it. The change happens; the UI just does not read it. Open a new session, come back, and you will see it was applied. Steering is hit or miss. Sometimes a steering message redirects the conversation right away. But most of the time the agent ignores it completely and waits until the whole task is done, or it spends a really long time on the task before the steering finally kicks in. I always wish steering was instant so no time gets wasted. Subagents are excellent, but sometimes Hermes hands me a result and then waits for the subagent's report, instead of waiting first and giving me one final result. Computer use is one of the most important features for me, and it sometimes breaks and needs a real fix. It happened enough that I stopped using it heavily. It usually breaks under heavy work or with medium and weaker models, so give it your most capable models. That brings me to the most dangerous problem. Weaker models can wreck your Hermes in minutes. You build your setup and the agent's understanding of you with a serious model, then a weak model starts making bad decisions and damages that whole layer. It can even break the app. MiniMax M3 broke my Hermes more than once with its hallucinations. The fix is simple: stick to the leading, stable models, such as Kimi, Sol, GLM 5.3, DeepSeek Flash, and frontier Qwen models like Qwen 3.8 Max, and stay away from weak, hallucinating ones. Since I stopped using MiniMax M3, I have not hit this once. I do not know if it has been fixed, but I am not risking my stability to find out. What I really want is an optional protection layer for the memory, the soul file and AGENTS.md, so any edit shows up in the interface for me to approve first. Long tasks can occasionally collapse in goal mode. The agent marks the task done before it is, or gets lost and loops. That happened a long time ago and has not happened since. It is rare and usually comes with weaker models, but I am mentioning it for transparency. One for Mac users: more than once, an update left my desktop app stuck on an older version. The fix is simple. Quit the app and run hermes update in the terminal. If the app still shows the old version, run hermes desktop to rebuild it. What I want next Two things matter most to me. First, a solid, easy mobile app with beautiful interfaces and that Nous spirit. Second, an easy mode that works out of the box: add profiles by field, so a researcher, a programmer, or a writer starts from a ready base instead of a raw zero. A model from Nous Research that understands Hermes better than any other model would be amazing too, but that is not central for me right now. And I want Hermes to switch modes from natural language. When I give it a goal, it should turn on goal mode by itself and finish the whole task, the way Codex and sometimes Claude Code do. The community and the team One moment made me appreciate Hermes more than ever. I wanted Arabic support in the harnesses I use, so I reached out to MiniMax, Zai and others and offered to translate the harness, the website, everything, myself. They all said yes. Then nothing happened. With Hermes, I did the Arabic localization myself and opened a pull request. They accepted it, and today Hermes speaks Arabic. Your voice as a user is what matters, and that is exactly why I love Hermes. Another thing that amazed me: one day I sent the team an idea right here on X. I wished I could set the model for each scheduled task. In less than three hours it was built and shipped, and I have been using it ever since. I have never seen that kind of speed. That is what accelerating growth looks like when a project is open source and the whole community builds it. Where I stand I have tried most of the well-known tools in the space, and every tool has beautiful positives and annoying negatives. Hermes might be the one with the fewest negatives. Not because it is better or worse than the others, but because in Hermes you can deal with a negative, change it, erase it, and sometimes turn it into a strength. It all comes down to how you customize it. It is not the easiest tool to start with, so my next post is my full setup, as a skill. Hand it to your agent, answer its questions, and it sets everything up for you. I do not claim it is the best setup, only the one that fits me. Stay tuned! After all this time, I know exactly where I stand. Hermes is a victory for what open source is supposed to mean.
115
84
22
941
79,919