CryptoPepper

vip
Age 3.4 Year
Peak Tier 2
No content yet
Sincerely recommend this GitHub
It's not written by me, but it's amazing
You can think of it as a unit test for the spec, development in 2026 is already in a spec-heavy, code-light mode
The spec is crucial to the life and death of your project, honestly I think specs also need access control and CICD
Hope there can be a framework developed, code has already become cheap, but specs are not
View Original
  • Reward
  • Comment
  • Repost
  • Share
Offshore RMB is at 6.8258, with an intraday range of 6.8213-6.8344
The trend of the RMB this year is completely different from the past few years.
The main theme from 2022 to 2024 is "Dollar strong + China weak = RMB depreciation". Now the main theme has changed—dollar peaks and falls back + China's economy stabilizes + capital flows back to emerging markets.
The 6.8 level is a key node.
Looking downward (RMB continues to appreciate), the driving forces come from: dollar weakening, China's export data consistently exceeding expectations, and northbound capital inflows.
Looking upward
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Just a few words: what I’ve been doing these days
1. Developing the MAC version of Hyper投OS (AI hackathon project). One-click to launch 8 AI Agents for your exclusive job-search ops (screening/applying/communication), covering the globe
2. A market maker trademark detector (via using ML to unlearn/learn away massive targets — learn the rules — learn the trends). First raise early warnings — run control groups — build strategies
3. Even driven + a market maker trademark detector — keep running control groups/token unlocks/expected listings scores. This is another track
4. Still event-dr
View Original
  • Reward
  • Comment
  • Repost
  • Share
Friends are long-term investors in U.S. stocks, and this year they increased their international allocation to 25-30%.
He says U.S. stock valuations are too high, the dollar has peaked, and some sectors in Europe and Asia have better growth prospects.
Since the beginning of the year, the trend of funds flowing out of U.S. stocks and into international markets has become very clear.
The market share of U.S. stocks in the global stock market has started to decline from its high point in 2024.
The MSCI EAFE (developed markets excluding the U.S.) has begun to outperform the S&P this year.
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
A battery recycling company that processed $1.1 billion has filed for bankruptcy.
Ascend Elements, one of the largest lithium battery recycling companies in the United States, filed for Chapter 11 bankruptcy protection on April 9. Liabilities exceed $192 million.
Founded in 2015, with more than $1.1 billion in funding (equity + government grants), it has a plant in Georgia that specializes in dismantling old batteries and reprocessing them into cathode materials.
Because
The Trump administration canceled a $316 million government grant. This funding was originally earmarked to build a
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Not just the United States
The Bank of Canada is also discussing AI cybersecurity risks with banks.
In the U.S., Bessent and Powell convened CEOs of major Wall Street firms. Almost simultaneously, the Bank of Canada held discussions with its leading banks about AI threats.
When these two signals are combined, it indicates
Global central banks are evaluating the same risk: the speed at which an AI model can discover zero-day vulnerabilities has surpassed the patching speed of human security teams.
Mythos has identified thousands of zero-day vulnerabilities within weeks, affecting all
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Anthropic says Mythos is too powerful to be made public. As a result, small models can also reproduce most of the achievements.
A security company called Aisle conducted an experiment: using small open-source models to reproduce the vulnerability discoveries publicly demonstrated in Mythos.
Anthropic's main reason for not releasing Mythos is "too dangerous, hackers could cause disasters if they get it." But if small models can do similar things, then the strategy of "not releasing" itself becomes untenable.
Aisle's exact words: models are just engines; the real attack/defense capabilitie
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Gemma 4 is finally stable on llama.cpp
On April 2nd, Google released Gemma 4, and on the first day, llama.cpp support was available but with many bugs. Now all issues are fixed
E2B, E4B, 26B MoE, 31B Dense
31B ranks third in Arena AI leaderboard, 26B ranks sixth
The strongest tier of open-source models
Use --chat-template-file to load interleaved templates
It is recommended to enable --cache-ram 2048
Context length depends on VRAM
Last year, the best local model was Llama 3.1 70B quantized version, barely usable
Now, Gemma 4 31B Q5 runs smoothly on Mac Studio, approaching GPT
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Stop arguing. The Jiangsu Alliance has been formed.
Only trusted acquaintances are being invited in.
"Jiangsu is a concept."
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
A startup company aims to break China's rare earth monopoly with a plasma reactor.
Radify Metals has raised less than $3 million, using plasma to process metal oxides to produce pure metals, with water vapor as the only waste.
Traditional rare earth processing requires大量化学试剂,污染严重,产线巨大。中国控制了全球70%以上的稀土加工产能,这是地缘政治里最硬的卡脖子
Radify's method: heat hydrogen to a plasma state, blow in metal oxide powder, strip away oxygen, and pure metal comes out the other end. The reactor is small in size and can be deployed in a distributed manner.
Currently working on samarium and dysprosium—two key rare ear
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Meta released Muse Spark, the first non-open-source model
Llama allowed Meta to take the lead in open-source AI discourse
Now Muse Spark is closed source
Alexandr Wang was recruited from Scale AI to do this, achieving competitive performance with one-tenth of the computing power
Meta has shifted from "letting everyone use it" to "only my users can use it." This is the first step toward Appleization; does Meta's stock say yes to this move?
View Original
  • Reward
  • Comment
  • Repost
  • Share
Someone on Reddit summarized Anthropic’s press release playbook: “We built something incredibly powerful. It’s so powerful that we can’t let you use it. But we want you to know just how powerful it is. Please compliment us for being responsible.”
Claude Mythos is the latest version of this template.
It has found vulnerabilities in every mainstream OS and browser, so it’s only used selectively by 40 companies.
Honestly, this really is the best PR strategy—“Too powerful to release” is 100 times more effective than “Come use it quickly.”
View Original
  • Reward
  • Comment
  • Repost
  • Share
A 23-year-old German kid raised $15 million in seed funding with a valuation of $200k.
He integrated an AI agent into iMessage.
The logic behind Poke is actually very simple — you don’t need to download any app, learn any new interface, just send a message in iMessage, and the AI will do the work for you. Reply to emails, pay bills, reschedule meetings, book flights.
Led by General Catalyst, with investments from Stripe, Dropbox, and OpenAI team members.
6,000 beta users within Silicon Valley VC circles, sending 200k messages per month.
A truly useful AI product doesn’t need to educa
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
I found a magic tool that solves my biggest headache in video creation.
Friends who create content know—you're shooting in portrait mode, but uploading to YouTube requires landscape. Cropping cuts off parts of the frame, adding black borders looks ugly.
Now there's an open-source tool that uses AI to extend the footage outward, truly generating new content that seamlessly matches the original scene.
This tool is called ComfyUI-Wan-VACE-Prep, based on Alibaba's Wan VACE model.
It's very easy to use: load the video → select extension direction → wait for generation. Key parameters:
The
View Original
  • Reward
  • Comment
  • Repost
  • Share
Read a short excerpt from CZ @cz_'s new book
1. Relates deeply to the experience on Van, especially the part about King Edward and Hasting
2. Surprised that these Wall Street hedge funds entered Bitcoin so early, such as Susquehanna, Jump Trading, Citadel, Optiver, Two Sigma, Five Rings Capital, etc.
BTC0.92%
View Original
  • Reward
  • Comment
  • Repost
  • Share
Nvidia-supported Firmus.
Raised $1.35 billion in 6 months.
Valuation of $5.5 billion.
An Australian AI data center company.
Funding speed matches SaaS unicorns.
The computing power gap is so large that capital is willing to bet on any team with GPUs.
You can't get GPUs, others will help you build the data center.
View Original
  • Reward
  • Comment
  • Repost
  • Share
Iran is gradually becoming more Trump-like
View Original
  • Reward
  • Comment
  • Repost
  • Share
In the future, please ask the local model to handle these token-consuming tasks.
Especially for something like autoresearch—only through multiple-iteration loops can it produce results.
1. As long as you have a spec and a clear goal
2. A complete structural framework + backtesting
Then you can use qwen 3.5 locally, running it 24/7 endlessly to churn out a strategy that satisfies me—feel what “capitalist exploitation” really means.
View Original
post-image
  • Reward
  • Comment
  • Repost
  • Share
Life is only taxes/death
Whatever you write/ tobacco/SM/crypto/AI are just the scenery along your journey
View Original
  • Reward
  • Comment
  • Repost
  • Share
  • Pin