WIll

🚀 Developer | 🤖co-founder of @valuecell

[WIll @Finsignals]

🚨 Tesla (TSLA) Update
Price: $340.94 (+0.27%)
High: $342.1 | Low: $335.35
📈 Boosted by:
Powell’s dovish tone → rate cut hopes
$7,500 EV tax credit extension
Progress on FSD v14 & Robotaxi
TSLA testing key $338 resistance — breakout could mark a turnaround.
#Tesla $TSLA #StockMarket #EV

[WIll @Finsignals]

OpenAI just signed a massive deal with Amazon: $50 billion in investment and a cloud contract worth over $100 billion over eight years. OpenAI's models are officially coming to AWS Bedrock

The real story here is how OpenAI is talking about Microsoft now. Their new revenue chief said in an internal memo that the Microsoft partnership "limited our ability to meet enterprises where they are."

Microsoft has poured $13 billion into this relationship, and the response is being publicly called a bottleneck. That tells you everything about where things stand

From OpenAI's perspective, this move was inevitable. A huge chunk of enterprise customers run on AWS. Locking your models behind Azure means handing market share to Anthropic and Google on a silver platter. Enterprise already makes up 40% of OpenAI's revenue — you can't keep growing if you're chained to one cloud

The smartest player in all of this is Amazon. They've invested $50 billion in OpenAI and $8 billion in Anthropic — the two strongest AI labs, both selling on their platform. They're not betting on which model wins. They're making sure every winner pays them rent. Same playbook as AWS in the early cloud days

Microsoft isn't sitting still either. They've already shipped their own MAI model family and put $10 billion into custom AI chips. On the surface it's still a partnership. Under the hood, both sides are building escape routes

Zoom out and the bigger picture is clear. AI model capabilities are converging — Stanford's latest report puts the gap between the top US and Chinese models at just 2.7 percentage points. When everyone's models are roughly the same, what decides the winner is who gets them into customers' hands faster

The technology race is over. The distribution war just started

[WIll @Finsignals]

GitHub: https://t.co/W1XKjZ9ZaB

[WIll @Finsignals]

NVIDIA just open-sourced PersonaPlex — a 7B speech-to-speech model that does two things really well: full-duplex conversation and persona control.

Full-duplex means you can interrupt the AI mid-sentence and it handles overlapping speech naturally. No more awkward "wait for the beep" turn-taking. It actually feels like talking to someone.

Persona control is the other big piece. You define a voice, a tone, a personality — and it stays consistent throughout the conversation. No character drift. No random style shifts halfway through a dialogue.

Why this matters:

→ 7B params, purpose-built for speech. Small enough to be practical, good enough to rank on speech-to-speech leaderboards

→ End-to-end real-time processing targeting low latency (related Nemotron models hit sub-300ms)

→ Clear use cases: game NPCs, virtual assistants, customer service — anywhere you need an AI voice with an actual identity

→ Community is already moving fast. ComfyUI nodes and an MLX port for Apple Silicon are already out there

Everyone's focused on text model benchmarks right now. But voice is the interface where AI actually meets everyday life. Full-duplex + persona consistency is what takes voice AI from "it works" to "it feels right."

[WIll @Finsignals]

🔗 https://t.co/55n2lWmghu

[WIll @Finsignals]

No content ideas for short videos? Try this open-source tool 👇

RedditVideoMakerBot — one command to automatically scrape trending Reddit posts, generate TTS voiceover, take screenshots, and edit everything into a ready-to-post short video.

TikTok / YouTube Shorts / Instagram Reels — works across all platforms.

The ultimate productivity hack for content creators. No CapCut, no Premiere Pro — just code.

[WIll @Finsignals]

Karpathy made an observation recently that really nails it: the AI world has split into two completely parallel universes.

On one side, the spectators.

They use free-tier ChatGPT, or watch AI fail compilations on short video apps, and conclude the whole thing is a joke. Ask it to write something, get a mediocre result, case closed — "bubble." Their entire understanding of AI comes from free tools, secondhand jokes, and screenshot hot takes.

On the other side, the actual users.

Claude Code and Cursor running all day, with them stepping back into the reviewer role. A lot of these people aren't just anxious anymore — they're genuinely shaken. You watch AI write an entire codebase from scratch, with cleaner logic than yours, faster than you, catching edge cases you didn't even think of — and then it hits you: the skills you spent four years of college and three years of work building, it just replicated in under an hour.

Karpathy calls this "AI Psychosis." I think a more accurate term is professional identity crisis.

So here's the real question: why does the same model code like a senior engineer but write like an intern?

The technical answer is straightforward — code has unit tests, clear right-or-wrong signals that models can learn from. Writing? There's no objective benchmark for "good." The model has no clean feedback loop to optimize against.

But there's another reason nobody talks about: money dictates priorities.

The biggest revenue stream for OpenAI and Anthropic isn't consumer subscriptions. It's enterprise APIs, developer tools, B2B solutions. Their best engineers and most compute are all poured into helping businesses write code, process data, and automate workflows.

Your writing experience? Your search experience? Sorry, not a priority.

A few uncomfortable truths to close:

If you've only used the free tier and already decided "AI is overhyped" — you haven't even seen what the best models can do. What exactly are you judging?

If you think "AI has nothing to do with me" — as long as 30%+ of your job is mechanical input-output tasks, you're already in the blast radius. You just don't feel it yet.

And if you're already using SOTA models and starting to panic — your fear is valid, but possibly misdirected. AI isn't replacing "programmers." It's replacing the "don't think, just execute" portion of the work. If your day is copying Jira tickets into CRUD code, yeah, worry. But if you're making architecture decisions, understanding the full business picture, balancing competing stakeholder interests — that stuff, AI still can't touch.

[WIll @Finsignals]

OpenAI just pitched investors a wild number: $2.5B in ChatGPT ad revenue for 2026, scaling to $100B by 2030.

6 weeks after launch, they're already at $100M annualized with 600+ advertisers. The growth is genuinely insane.

But here's what bothers me:

1/ Sam Altman was adamant in 2024 that OpenAI wouldn't go the ad route. Took less than two years to break that promise. Faster than their model release cycle.

2/ With search ads, you know what's an ad. But in an AI conversation? When ChatGPT says "I'd recommend trying this product" in first person, can you really tell if that's genuine reasoning or a paid placement? The line between advice and advertisement has never been blurrier.

3/ Free and Go ($8/mo) users now see ads. Want a clean experience? Pay $20. Classic "degrade the free tier to push upgrades" playbook. The free users aren't the customers — they're the product.

4/ What really worries me isn't today. It's what $100B in ad revenue by 2030 demands. To hit that number, ads will creep from small "Sponsored" cards into every corner of every conversation. At that point, is ChatGPT still your AI assistant, or just a billboard that talks back?

We've seen this movie before. Google went from "Don't Be Evil" to four ads above every search result. Took about 15 years.

OpenAI is speedrunning that arc at 10x.

The one silver lining: users have choices this time. Claude, Perplexity, open-source models — AI isn't winner-take-all like search was.

Let's hope competition keeps them honest. But I wouldn't bet on it.

[WIll @Finsignals]

Install it, ask it anything about negotiation or power dynamics, and get answers running on Trump logic.

👉 https://t.co/dam8MsmBfk

[WIll @Finsignals]

What if Trump's negotiation brain was reverse-engineered into an AI skill?

Not an impression. Not "talk like Trump." The actual operating system underneath — extracted from 30+ primary sources, distilled into 6 mental models and 8 decision heuristics.

Everything is a deal. Open high. Be unpredictable. Control the narrative. Never admit defeat. Know when to fold — and make it look like a win.

[WIll @Finsignals]

GitHub: https://t.co/jZA616J13j
Website: https://t.co/Ye0F7P0wC7

[WIll @Finsignals]

Stock research is painful.
You read news in one app, track positions in another, save ideas in spreadsheets, and still end up with a messy decision process.
Found ValueCell recently, an open-source AI investing platform built to make portfolio research and stock analysis more organized.

It brings together:
watchlists and portfolio tracking
AI-assisted stock analysis
market news and signals
a more structured workflow for investment research

What I like is that it’s not just another dashboard full of numbers.
It tries to turn scattered information into a research process you can actually use.
If you care about AI for investing, open-source fintech, or building a smarter stock research workflow, take a look
#Github #opensource #valuecell #investing

[WIll @Finsignals]

everything-claude-code was open-sourced just a few days ago, and it’s already at 120K+ stars on GitHub.

The author shared their full Claude Code setup: agents, skills, slash commands, rules, MCPs — the whole workflow.

And this isn’t just “nice-looking config.” It’s packed with things you can actually use:

how to save tokens, choose models, and slim down system prompts
how to give AI memory and continue work across sessions
how to make it summarize experience from conversations and improve over time
how to evaluate outputs, score results, and make them more reliable
when to run tasks in parallel
how to coordinate multiple agents without chaos
A lot of the hard parts are already figured out: the scripts are organized, the docs are solid, and it works on both Windows and Mac.

Even better: this setup wasn’t built for show. The author used it in a hackathon — and won first place.

If you use Claude or Cursor regularly, this is probably worth stealing from instead of rebuilding everything yourself.

[WIll @Finsignals]

Today’s top trending GitHub repo: luongnv89/claude-howto

A visual, example-driven guide to Claude Code — covering everything from core concepts to advanced agent workflows, with practical copy-paste templates you can use right away.

Why developers are watching it:
Clear, hands-on examples
Useful for both beginners and advanced users
Focused on real-world agent workflows, not just theory

📈 4,232 stars today
🌟 10.8K stars total
Repo: https://t.co/ul4NWgCdvM

#GitHub #OpenSource #AI #ClaudeCode #AIAgents #Developers

[WIll @Finsignals]

This is pretty cool 👀

ValueCell is building an open-source multi-agent platform for finance — covering stock research, market analysis, news tracking, and strategy backtesting.

If you’re interested in the future of AI-native investing tools, worth a look.

#GitHub #opensource #AI

[WIll @Finsignals]

https://t.co/aPvb17pxWw

[WIll @Finsignals]

Breaking the Information Barrier: How Pi + E2B @e2b Empowers ValueCell @valuecell Beyond Traditional AI

For many, the perception of AI is still stuck on the idea of a "smarter chatbot." Consequently, traditional AI constantly spins its wheels within the confines of "known information." But the true ceiling of an Agent is rarely determined by its conversational fluency; it is defined by its ability to drive tasks to completion when information is incomplete, data is missing, and the real world is highly dynamic.

This is exactly where traditional AI hits a wall:

Knowledge cutoffs resulting in a lack of real-time awareness.

Incomplete retrieved information leading to ungrounded, hollow conclusions.

When facing data gaps, it resorts to "guessing an approximate answer."

It can analyze, but it cannot take physical action to fetch and fill in the missing data.

The core issue has never just been "is the model powerful enough?" It is this: When existing information falls short, does the AI possess the agency to actively push the boundary outward? This is the exact bottleneck ValueCell is engineered to shatter.

Not a Solo Agent, but a Dynamic Collaborative Graph
ValueCell's underlying engine isn't a simple point-to-point model hooked up to external tools. It functions much more like a Multi-Agent dynamic Graph built on ReAct principles. The core Planner first deciphers user intent, then breaks the task down into executable sub-problems. From there, it:

Prioritizes reusing existing skills.

Dispatches appropriate dedicated agents based on the specific problem type.

Triggers different agents via natural language to handle routine tasks.

Only when standard capabilities fall short does it hand the problem over to a highly flexible coding agent.

In real-world business environments, a massive volume of problems doesn't require a coding agent to write scripts from scratch every time. What can be standardized must be standardized; what can be reused must be reused.

Conversely, long-tail problems will always exist. And the higher the value of the problem, the more "long-tail" it tends to be. Relying solely on a fixed combination of skills can never cover the totality of real-world tasks. The system ultimately demands a sufficiently agile coding agent to tackle issues outside predefined workflows.

The True Value of Pi: Conquering the Long Tail
Within ValueCell, Pi operates as the coding agent designed for complex, long-tail scenarios. It steps in to handle open-ended, irregular tasks when existing skills and agent orchestration cannot directly resolve the issue. For example, when the system needs to:

Scrape data from a public webpage on the fly.

Sanitize and unify multi-source unstructured information into standardized fields.

Write a script to validate a specific hypothesis.

Feed execution results back into the upstream pipeline for continued reasoning.

This is where Pi's edge becomes apparent. Compared to other coding agent providers on the market (like Claude Code or Codex), ValueCell's selection of Pi isn't based on empty claims of being "the absolute strongest." It was chosen because of critical engineering traits proven in practice:

Exceptional flexibility.

High fault tolerance and permissiveness.

Strong observability.

These characteristics are fundamentally more critical for production environments than simply posting impressive benchmark scores on a single-turn demo.

E2B: More Than Just a Code Interpreter
If Pi solves how to flexibly write and invoke code, E2B solves the equally critical engineering problem of where these executions should occur.

ValueCell integrates a sandbox not just to tack on a "cool" code interpreter feature, but out of absolute necessity for service stability, environmental isolation, data security, and execution control. The sandbox is a fundamental engineering architecture choice, not a gimmick. It ensures the coding agent can securely execute within an isolated environment to handle:

Script execution

Web scraping

Data sanitization

Structured processing

Result generation and feedback

This grants the system agile execution capabilities without dragging the main service into a volatile runtime environment.

The Critical Differentiator: Forging a Closed Loop
Many systems have integrated sandboxes, but most rely on one-way calls: Main System Dispatches Task → Sandbox Executes → Returns Result.

While this pipeline functions, its limitations are glaring—it remains a one-off, unidirectional tool invocation. ValueCell's architectural breakthrough is that it doesn't just throw the coding agent into a sandbox to run in isolation. Instead, it constructs a secure, bidirectional collaboration mechanism:

The Planner orchestrates the coding agent.

The coding agent enters the sandbox to execute.

The execution process inside the sandbox forms a continuous loop back to the system via secure channels.

Execution results, context, and data capabilities are not disjointed; they continuously synergize. This means the coding agent's capability is no longer just "running scripts in an external environment"—it is genuinely integrated into the entire Graph's decision-making and execution loop. It guarantees the coding agent is a core system capability, not a bolted-on plugin.

The Paradigm Shift: Pushing the Information Boundary
Ultimately, what Pi + E2B pushes open is not just the boundary of tooling, but the boundary of information itself.

The flaw in traditional AI isn't an inability to answer; it's an over-reliance on pre-existing information. ValueCell's architecture ensures that when the system encounters an information gap, it doesn't halt or guess blindly. Instead, the Planner assesses the problem, routes standard tasks to dedicated agents, and escalates unresolved long-tail problems to Pi. E2B provides the secure execution environment, and the results are safely fed back into the system to drive the task forward.

The entire paradigm shifts from "organizing an answer based on what is known" to "actively acquiring new facts mid-task, bridging the gaps, and continuing the reasoning process."

It dominates the scenarios where traditional AI crashes: missing incremental data, fragmented sources, unusable unstructured content, and complex workflows requiring mid-flight correction. Where traditional AI can only "guess," ValueCell physically steps in to scrape, clean, compute, and fill the voids. Furthermore, ValueCell's coding agent doesn't just write code to backfill data; it instantly generates interactive ECharts, seamlessly stringing together extraction, processing, analysis, and visualization with maximum agility.

The next era of Agent competition won't be won by "who speaks better." It will be won by: Who can persistently drive a problem to completion in an information-incomplete world.

From this perspective, the integration of Pi + E2B is never just about "making AI write code." It is about equipping the Agents within ValueCell with the definitive execution power to systematically push the boundaries of information ever outward.

[WIll @Finsignals]

🚀 Tired of quitting OpenClaw because of confusing setups?
Meet Claw X — the free, open-source desktop app that makes OpenClaw actually usable.

✨ What you get:

One-click install via setup wizard — no terminal, no commands, no configs

Telegram, Discord, WhatsApp all in one dashboard

Schedule tasks, manage skills, swap AI providers — all visually

Works with Anthropic, OpenAI, Google, even local models via Ollama

Or let OpenClaw install it for you with a GitHub link

#OpenClaw is insanely powerful — #ClawX finally makes it accessible.

[WIll @Finsignals]

We built https://t.co/H0L3K5vsE8 as a chat-based financial research tool — you ask market questions, AI agents go find the answers.

But instead of explaining how it works, let me just show you what people are actually asking on the platform today:

"which sectors had the strongest fund inflows this morning?"
"NVDA crushed earnings — what stocks benefit from their supply chain?"
"fiber optic stocks all hit limit-up today. real demand or pure hype?"

The difference from just asking ChatGPT the same questions:

→ https://t.co/H0L3K5vsE8 pulls live market data. not training data from 6 months ago.
→ when you ask a question, multiple agents work on it — one handles financials, one tracks news, one checks risk. you get a layered answer, not a single model guessing.
→ it covers US, HK, China A-shares, and crypto. one conversation, all markets.
→ your chat history is saved, so you can build on yesterday's research today.

A few ways people are using it daily:

• post-market review — "summarize today's fund flows and sector moves"
• earnings analysis — drop a ticker after earnings and get a breakdown vs expectations
• cross-market reads — "how does nvidia's guidance affect A-share optical module stocks?"
• sanity checks — "i'm bullish on X, tell me what i might be missing"

Free to try, no setup needed. just go to https://t.co/H0L3K5vsE8 and start asking.

Honestly the best way to get it is to just try a question. pick something you were going to research today anyway and see what comes back.

https://t.co/H0L3K5vsE8 🧪

[WIll @Finsignals]

If you're still struggling to build a financial research system with OpenClaw, we're bringing OpenClaw's power directly into ValueCell.

No more wrestling with deployments, skills, or command lines. Just open the app and start analyzing.

What's happening:

We're embedding OpenClaw's core capabilities — code generation, data scraping, automated workflows — natively into ValueCell's investment research platform.

This means you get: → OpenClaw writing custom scrapers for you behind the scenes → Automated strategy backtesting without touching a terminal → Multi-agent research pipelines that just work out of the box → A/US/HK stocks + crypto, all in one place

Why we're doing this:

We kept seeing the same pattern. People excited about OpenClaw for finance. People spending days setting up environments, writing skills, debugging API connections. People giving up before they ever analyzed a single stock.

The tool was getting in the way of the work.

So instead of asking analysts to become DevOps engineers, we're packaging the engine inside the car.

OpenClaw's flexibility is incredible for builders. But if your job is researching markets, not configuring agents, you shouldn't have to earn a CS degree first.

ValueCell + OpenClaw = the research depth of a custom agent setup, with the simplicity of a product that respects your time.

Shipping soon. More details to come.
#OpenClaw #valuecell #AI

[WIll @Finsignals]

The Real Skill in 2026

GitHub Copilot: 1M+ developers, 3 billion lines of code generated, 55.8% faster task completion.

When code generation approaches zero marginal cost, what's actually valuable?

Not writing code. Knowing what to build.

→ Understanding real user pain → Designing the right interaction → Deciding what NOT to build → Knowing when to ship

Every vibe coder can spin up an app in 2 hours now. But most of those apps solve problems nobody has.

The 2026 competitive advantage isn't coding speed. It's thinking quality.

AI gave everyone a brush. It didn't give everyone taste.