PΞTE | Builder, Trader (危機)

Day 106: Building the first Vibe Charting Platform This is the one I've been waiting for. 🚀 The moat. The AI stuff. Finally here. 🏁 What the agent can do right now: ✅Workspace-aware — knows your open charts, symbols, timeframes ✅Harness tool to feed chart data into the conversation — bars, indicators, drawings ✅Draws lines on price levels directly on your charts ✅Marks price ranges with rectangles on your charts Next in backlog, agents should be able to: 🔧Fully build setups and backtest them automatically 🔧Run multi-timeframe analysis across open charts 🔧Learn user-scoped SKILLS — your way of trading If you want to test the app, shoot me a DM or drop a comment below!

You are the only one who knows what "Double" means. Double from yesterday's limit from 2 weeks ago? There is zero transparency for us. And even with "double", your limits are still a joke. Been a max sub for 11 months and last 3 have been lame, I'm pissed on me I didn't switch faster. Zero customer ethics!

Effective today, we are: 1) Doubling Claude Code’s 5-hour rate limits for Pro, Max, and Team plans; 2) Removing the peak hours limit reduction on Claude Code for Pro and Max plans; and 3) Substantially raising our API rate limits for Opus models.

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.

I switched a few days ago to codex after using claude max for 11 months and I'm impressed. The limits barely budge, I know it's 2x this May but even without this bonus it's probably better by an order of magnitude. And the model is very very capable. However, the harness functionalities that I miss the most are: - /rc command - starting a session from my pc and monitoring it from my phone while I run errands or do house chores, I've got pretty used to having that. - ultrathink - I noticed the model has very good reasoning on high and even on medium. For the most complex tasks I'd like to trigger xHigh without having to change the reasoning. - Agent Teams: for vertical and bigger story implementations (i.e.: story that touches 2 microservices + api gateway + FE), I relied heavily on agent teams where the work is broken down between agents that can communicate with each other, and can create tasks for each other, If BE agent has to change an endpoint signature it can notify the FE agent etc. This can't be done with the current subagent orchestration. This gave me great peace of mind as none of the agents in the team got to the point of compacting their context window. Speaking of compacting, I try not get to that point too often, but I noticed codex compacting is way better than claude's which feels like it goes out of it's way to lose all the important information and keep only nonsense.

What are we obviously not getting right with Codex?

Yes, last cycle felt very sad for me personally. I now feel the same excitement I felt in reading whitepapers and playing with defi protocols, by tinkering with things like OpenCalw and Hermes. Too bad we can't have any skin in the game. Memes took away the attention from crypto projects like that.

@pete3m You’re describing a real sentiment a lot of people in crypto share right now: the feeling that the “retail-driven innovation + altcoin cycle energy” has faded compared to 2017 or even 2021–22.

Most of the crypto industry is dead and never coming back. Eventually people will realize it.

I don't think that was the issue. In 2017 I invested in dozens of ICOs, you could enter in private / seed rounds pretty easily as well. In 2022 it got way harder for the retail, the majority of tokens were sold only to VCs leaving a few percentage of the supply for retail as "marketing budget". And to add fuel to the fire, VCs we're dumping on retail at TGE, so meme coins came as a natural response. I think the biggest contributor to this was pump fun not Elon.

@pete3m @APompliano Elon killed crypto the minute he started talking about doge coin. Then meme coins exploded.

Strongly disagree. But yes, where money runs around scammers will exist, that's regardless of industry/domain. And yes very few projects added real value, just like very few remained after .com bubble. Giving the little guy a chance to invest and grow was a too good to be true value proposition. Try getting into AI deals now, see how hard / next to impossible is. By the time they IPO, the "insiders" will have extracted massive growth, and retail is back at holding the bag as exit liquidity.

@pete3m @APompliano There never was any innovation. Bitcoin was the innovation. Bunch of vaporware and scams. The culture was get rich quick and sell to the next biggest idiot

@yantsab Inconsistency in the reasoning level and/or limits. The limits for the last 2-3 months have been an absolute joke.

@pete3m What was the trigger?

After 11 months of using Claude Code, I switched to Codex. I guess I'm a codex guy now...

@cryptopunk7213 The thing that bothers me the most is the lack of transparency. They dance around diminishing compute and diminishing limits. This makes it extremely frustrating, you pay a fixed amount each month and get back "some" AI.

@pete3m i hear you and have been seeing the same. i hope they lock in more compute to serve these models

new opus 4.7 model appears to be a hybrid of mythos, it actually beats it in some cases: - almost every benchmark is within 50% range of mythos. - beats mythos at financial engineering (very good for enterprise use) - convincingly better at coding than gpt 5.4 and gemini (11 point increase) - worse at browser use than opus - great visual reasoning (reading images), works well on long range tasks. seems like a dope model, wish it was mythos tho :)

@ustas_eth @claudeai I'd whish! It's more like Sonet 3.5 benchmarks by next Tuesday. What a joke...

@claudeai fixed for you

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

@MarceloRet41877 @claudeai @AnthropicAI Anthropic has absolute ZERO client ethics!

@claudeai @AnthropicAI They degraded 4.6 and now delivering 4.7 with the same quality as when 4.6 was at its peak. People will forget about it and call it awesome.

@claudeai with all the nerfing you've been doing lately, those numbers mean nothing!

@AnichoBuilds That's interesting, I'll try doing some tests.

@pete3m Genuinely found it's.performing better via cursor or other harnesses than Claude code cli

Is it me or is max effort Opus 4.6 degraded to the level of a Sonnet 3.5? It's like it went full retard all of the sudden! Cannot handle super simple prompts with empty context. Drives me nuts. How can it fluctuate like this...

Day 110: Building the first Vibe Charting Platform Haven't posted my progress for the last few days Working on some boring stuff 🥱, like: - adding webhooks to keycloak for syncing new users with my platform - added backup worker service in case webhooks fail - added custom theme to keycloak auth screens - enabled Google OAuth for easy signup/login.

@gecko002002 Wrote in caps all day today... geez

@pete3m 100% 😅😅

@Austen Has been a mess lately...

To be fair to Anthropic/Claude: Can you even imagine how much scaling they have to do? They’re adding 1 million users/day at the same time as they’re releasing new features daily, and users can absolutely gobble compute. It’s a miracle these AI services are up at all.

@JoshKale Wen Mythos? Let's go!

Anthropic just accidentally leaked the most powerful AI model ever built and even more stocks are getting CRUSHED. A CMS misconfiguration exposed 3,000 unpublished files, including a full draft blog post announcing "Claude Mythos," a model so advanced that Anthropic created an entirely new tier for it. They're calling it Capybara. It sits above Opus as a fourth tier that didn't exist until now. Here's what the leaked documents reveal: - Mythos scores "dramatically higher" than Claude Opus 4.6 on software coding, academic reasoning, and cybersecurity - Opus 4.6 had just topped Terminal-Bench 2.0 at 65.4%, beating GPT-5.3-Codex. Mythos BLOWS past it. - On cybersecurity, Anthropic says Mythos is "currently far ahead of any other AI model in cyber capabilities" - They warn it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders" The company that built the model is telling us that defenders can't keep up and market believed them: - CrowdStrike: -7.7% - Palo Alto Networks: -7.5% - Zscaler: -8.4% - Okta: -7.3% Anthropic isn't planning a public release. The model is too powerful and too expensive. Early access goes only to organizations focused on cyber defense. They're arming the good guys first. The irony of the world's most advanced cybersecurity AI being exposed by a misconfigured content management system is almost too perfect. We're entering a phase where the most capable AI models may never be released publicly. Mythos is the first model built for a world where AI breaks software faster than humans can fix it

Day 109: Building the first Vibe Charting Platform Finally! No more hardcoded models all over the place! ✅UI for importing models from Open Router ✅Daily Chron job model syncing ✅Reflection selects all agents from code at runtime ✅Full CRUD UI for assigning models to agents.