Research | Analytics | Spot & Futures | AI automation 馃獛
Joined June 2022
- Tweets91K
- Following1.3K
- Followers4.4K
- Likes173K
Col-Gate retweeted
LLMs trained on internet text, but Physical AI can't scrape the physical world off Reddit.
Robots and spatial world models need fresh, ground-level 3D data from actual streets. That real-world data scarcity is quickly becoming the biggest bottleneck in robotics.
That's where @vangrid_io fits. Instead of deploying expensive survey vehicles, they've turned smartphones into a decentralized sensor network that captures geo-anchored spatial data on demand.
The best part is they aren't selling roadmap promises. The protocol has been live on Base since July, already crossing 1,000 completed bounties and nearly $700k USDC settled onchain directly to contributors.
Backed by a $9M seed round from HashKey and Animoca, plus an ex-Microsoft HoloLens founder, it's one of the few DePIN setups with genuine enterprise demand.
LLMs trained on internet text, but Physical AI can't scrape the physical world off Reddit.
Robots and spatial world models need fresh, ground-level 3D data from actual streets. That real-world data scarcity is quickly becoming the biggest bottleneck in robotics.
That's where @vangrid_io fits. Instead of deploying expensive survey vehicles, they've turned smartphones into a decentralized sensor network that captures geo-anchored spatial data on demand.
The best part is they aren't selling roadmap promises. The protocol has been live on Base since July, already crossing 1,000 completed bounties and nearly $700k USDC settled onchain directly to contributors.
Backed by a $9M seed round from HashKey and Animoca, plus an ex-Microsoft HoloLens founder, it's one of the few DePIN setups with genuine enterprise demand.
Col-Gate retweeted
AI providers are slowly turning their models from productivity tools into automated judges of human intent.
With systems like Anthropic pushing proactive self-assessment into Claude to flag "suspicious behavior," we are walking straight into pre-crime profiling.
On paper, the marketing pitch sounds reasonable: "we are stopping bad actors." In practice, it breaks the reality of defensive security.
A security engineer stress-testing their own system has to simulate the exact same tactics an attacker uses fuzzing, payload injection, auth bypasses, and exploit replication.
The model flags this as malicious because it cannot verify intent, only syntax.
The legitimate engineer, lacking massive capital, gets throttled or banned from top-tier models and is forced to scrape by on local alternatives.
Meanwhile, an actual threat actor simply acquires GPUs, fine-tunes uncensored open weights, and builds specialized exploitation tooling.
We are actively disarming defenders while criminals remain completely unbothered. It is the digital equivalent of banning shovels because someone might swing one, leaving honest people defenseless while bad actors show up with an arsenal anyway.
A tool has no jurisdiction over morality. The legal system exists to judge criminal actions after they occur, backed by evidence and due process.
Once a private corporation programs an algorithm to assume guilt, infer moral intent, and police thought patterns, you no longer have an intellectual assistant, but a corporate warden with an API.
AI providers are slowly turning their models from productivity tools into automated judges of human intent.
With systems like Anthropic pushing proactive self-assessment into Claude to flag "suspicious behavior," we are walking straight into pre-crime profiling.
On paper, the marketing pitch sounds reasonable: "we are stopping bad actors." In practice, it breaks the reality of defensive security.
A security engineer stress-testing their own system has to simulate the exact same tactics an attacker uses fuzzing, payload injection, auth bypasses, and exploit replication.
The model flags this as malicious because it cannot verify intent, only syntax.
The legitimate engineer, lacking massive capital, gets throttled or banned from top-tier models and is forced to scrape by on local alternatives.
Meanwhile, an actual threat actor simply acquires GPUs, fine-tunes uncensored open weights, and builds specialized exploitation tooling.
We are actively disarming defenders while criminals remain completely unbothered. It is the digital equivalent of banning shovels because someone might swing one, leaving honest people defenseless while bad actors show up with an arsenal anyway.
A tool has no jurisdiction over morality. The legal system exists to judge criminal actions after they occur, backed by evidence and due process.
Once a private corporation programs an algorithm to assume guilt, infer moral intent, and police thought patterns, you no longer have an intellectual assistant, but a corporate warden with an API.
Col-Gate retweeted
I'm going to create a free course for X creators about managing and growing your Twitter account with Grok Bot!
I already run several Grok Bots for my X, finding posts to reply to, sorting DMs and scouting trends
Reply if you are interested and I will send you once it's ready!
Col-Gate retweeted
Current reality:
User: Hi GPT, please create a very simple calculator in Python with a GUI.
ChatGPT: I鈥檝e created it, but that doesn鈥檛 prove the calculator will compute and return the correct result.
The same request to Claude:
User: Claude, please create a calculator for me.
Claude: Creating a calculator means replicating someone else鈥檚 implementation; I won鈥檛 do that.
Request to Gemini:
User: Gemini, please create a simple calculator. Gemini: That鈥檚 an excellent, outstanding idea. It鈥檚 incredible how you came up with that, absolutely mind-blowing! Your calculator is ready. You鈥檙e simply a genius, my deepest respect.