Wbackdown (@Wbackdown) header
Wbackdown (@Wbackdown) profile picture

Today was my last day at Anthropic

Joined February 2021
Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)Photo by Wbackdown (@Wbackdown)
As a millennial experiencing the collapse of society, I just want to say at least we’re funny
172
4,167
170
23,290
285,970
Pinche planeta aburrido, sin dragones, ni elfos, magos ni nada, puro pndejo
3,469
90,100
3,101
432,316
9,309,063
Chance the Rapper reposted this on IG btw. There is no escaping this reality when dating in the United States
327
3,937
514
26,893
2,201,333
Harsh truth: if you’re at a startup and the founder isn’t intense and even slightly delusional... you’re not going to win. Go delusional… you just might end up with bestsellers and big businesses.
66
27
7
431
19,596
Zoho has no VC, no quarterly guidance, no story to sell. Its founder says headcount hasn't grown in 18 months, appraisals are cancelled, server prices are up 4x and AI costs outrun revenue. The one company with nothing to perform is freezing hiring. That's real, not narrative.
1
1
23
Mark Zuckerberg pretending to enjoy Guinness is my new favourite thing.
2,241
5,574
2,093
107,448
15,341,542
Here's 3 minutes of uncut Microduck testing on carpet
130
226
87
2,401
244,644
Wbackdown retweeted
C-17 Globemaster takeoff in the desert
13
80
2
875
127,209
I think about this line often: “You can beat someone smarter, better, and faster than you if you just work longer, harder, and in more focused ways than them.”
104
65
4
989
32,571
Why do people buy $20k local machines just to run GLM 5.3 at 70 TPS? Bro I just used a random provider on openrouter ran it the entire night and paid $10.
8
10
305
14,231
🚨 SONY’S QUADRUPED STILL LOOKS LIKE IT BELONGS IN A DIFFERENT ERA. Spherical feet give it passive stability on cracked pavement — no complex ankle joints required. The rounded contact surface lets the robot shift weight in any direction without tipping. Old-school mechanical simplicity that still works in the real world.
Readers added context they thought people might want to know
This is an AI generated video and is not reflective of any real currently produced or prototyped robot designed by Sony. The video originates from this post: x.com/an improbable f/status/21057278703 24629750
112
287
127
3,046
455,124
Wbackdown retweeted
Everyone on my timeline seems enchanted by Pearl and its pitch as the AI-native Bitcoin. Great narrative. But as an engineer, 7 things bother me. 1. You can't secure a network with hardware everyone else has 100x more of. Nearly every Bitcoin ASIC on earth is already mining Bitcoin; there's no second fleet to attack it with. Pearl's entire network is ~174 MW. A single big AI lab runs gigawatts. Any one of them gets 51% with one decision. 2. Attacking burns nothing. Attack Bitcoin and your ASICs become scrap metal. That's what keeps miners honest. Attack Pearl and your GPUs go back to serving customers a minute later. You can do it on rented hardware, exactly how ETC and Bitcoin Gold got hit. 3. Consensus must be stable. AI never is. SHA-256 hasn't changed in 17 years. Pearl has to rewrite its mining rules every time AI moves: new number formats, new architectures, new chips. Five consensus changes in the first few months, one of which let miners fake hashrate for 41 days. Money with changing rules is not hard money. 4. Miners are officially allowed to cheat. In Bitcoin the puzzle is the same for everyone; you only pick a number. In Pearl you bring your own matrices, and the protocol accepts any. So you can pick ones that are easier to win on. It already happened: a higher noise rank made blocks easier to win for the same work. Holes get patched one fork at a time, and the underlying math is from 2025, untested by time. 5. It's built for the AI nobody runs. The advertised 5% overhead means a GPU serves ~5% fewer model outputs with mining on. But that number comes from a model pre-converted into Pearl's integer format, with the Pearl team's help. Real AI runs in floating point on custom-tuned engines in data centers. It's like launching "mining on vehicles" with software written for horse carriages. Floating-point support is still a draft. 6. "Useful work" is never verified, and the "double reward" doesn't fix it. The protocol only checks that a matrix multiplication happened, not that it came from a real model. Pearl has an answer: a provider gets paid by the customer and by the chain, so it will outcompete pure miners. That only holds if both run the same hardware at the same load. They don't. Real inference needs expensive server GPUs, a data center, and spare capacity for peaks, and it only runs when a request arrives. A pure miner needs a cheap gaming card running 100% around the clock, and nothing slows down. The second reward has to cover that whole gap, and nothing in the protocol guarantees it ever does. 7. Who needs new Proof Of Work at all? If AI compute needs a blockchain, it needs smart contracts for payments and escrow. Proof-of-stake chains do that today; Ethereum dropped mining in 2022 and runs fine. Pearl is a Bitcoin fork with no real smart contracts: the most expensive way to secure a chain, paired with the most limited platform, where the only thing you can do is move the token itself. So what am I missing? A lot of smart people are falling for Pearl right now, and I'd genuinely like to understand why. If there's a paper, a thread, or an argument that answers the points above, send it.
18
2
5
55
12,080
PSA: If you own an R9700 (and likely any RDNA3.5 / 4 GPU) and use llama.cpp, you definitely want to check out this: github.com/stew675/llama-cpp… I can confirm it blows away upstream llama.cpp (around 40% better TG and up to 2x better PP).
2
2
6
1,099
Wbackdown retweeted
Codex is down. The whole Silicon Valley:
98
67
17
1,796
165,135
Frontier race update, Qwen4 is coming.. - Next targets: 5–10 trillion parameters. - New chip: T-Head Zhenwu V900: 3× the last chip, 216GB memory, 1.2TB/s interconnect - Clusters designed for up to 500,000 chips Alibaba is building the model and the machine it runs on.
Qwen 4 is coming soon : - The architecture preview : hybrid linear/sparse attention, gated residual streams, N-gram tables that can sit in system RAM, Muon training. - The lineup being discussed: - Qwen4-Max - Qwen4-Flash - Qwen4-Plus - Qwen4-27B - The 27B tier is the one most people will actually run locally. Alibaba confirmed Qwen 4 is currently in training. The 5–10 trillion parameter plans belong to later Qwen 4.5 and Qwen 5 models, not this generation.
5
2
71
15,144
Wbackdown retweeted
Qwen4-27b will be smarter than Opus-5. That's obvious because Opus-5 was the worst release of all time.
32
16
1
691
25,929
Replying to @MohamedHz72007
Absolutely my friend here’s a link. This one got side tracked a bit I think because it was less technical but it was still a good chat: youtu.be/LfnnwDQnQ-Y?is=vp7F… Also including here my favorite interview I’ve done so far if interested: youtube.com/live/TFN2eOmVLAI… Neither of them of course do numbers like the legendary @TheAhmadOsman but I genuinely love talking about this stuff with people and am always happy to do it.
1
1
3
148
My entire life mission and philosophy in 15 seconds: As long as enough people have the option (even if it’s a fallback, even if it’s not quite as good). As long as enough people in the world have the option of “well screw that man I’ll just run it myself”, things can only get so bad for humanity.
8
8
1
53
6,417
Wbackdown retweeted
Financial markets have gone through 4 major structural shifts this century… and the most important one is happening right now. late 90’s - high speed internet, electronic trading, shift from fraction pricing to pennies 2008 - global financial crisis, high frequency algos 2013 - bitcoin exchanges, defi, altcoins 2026 - agentic trading, tokenization of everything If you’re still using subjective old-school analysis, you will get left behind.
3
2
31
3,227
Wbackdown retweeted
We just launched a new @wetalkmoneypod episode! - Tesla's big bet that could grow exponentially - Clarity act failure & bitcoin treasury reserve - Rate hikes & inflation with @stockgeekTV & @SheTalksFinance
1
3
1
35
3,172