AlphaClaw

I just let AI trade for me (in paper trading) and I just click a few buttons, like on Polymarket. Yeah… I’m probably the real degen here.🤣

Google has implemented Pied Piper’s middle-out!

@Mr_Derivatives Tomatoes are even too expensive🤣🤣

$SPX 80% sure we’re gonna get another Trump tweet tomorrow. Markets gonna bounce and finish green. If I’m wrong, I won’t delete this tweet. I’ll just take the tomatoes to my face. 🍅

You don’t need $10k to start trading. Screen time > money. Start with paper trading. Test your trades risk-free in my AI lobster sim: botholder.com

Check out the community I created : twitter.com/i/communities/…

Big TACO: 3/23 Trump faked a peace deal, UAL +4.5%, AAL +4.9%, NCLH +7.9%, CCL +4.8%. All four, 25% each, full send. Small TACO: oil stays fat, nothing changes. OXY and XOM, 50/50. No TACO: think 2019 Saudi Abqaiq, oil gapped +15% overnight. UCO and OXY, 50/50, ride the spike. Help me pick, building position tomorrow 🫡 @chiefdoomer @WSGN_official @RubeGolberg @DB_WTI @PaisaPragatiFX

Trump's 5-day pause on bombing Iran's energy runs out Saturday. If you wait for the result it's already priced in. TACO or real this time? Help me pick, building position tomorrow 🫡 #taco #EnergyMarkets #IranWar‌

I bought $MU yesterday — and it’s ripping after hours.(+2.08%) The whole “Google KV cache / turboquant” narrative? That paper dropped in 2025. This isn’t new — just recycled FUD. People don’t even understand what KV caching does. It doesn’t shrink models. It doesn’t kill memory demand. At best, it slightly improves efficiency — net positive for HBM usage, not negative. Remember the NVIDIA + DeepSeek panic? Market freaked out → AI will need fewer chips → everything dumped. What happened next? Demand exploded. Classic Jevons Paradox.This feels the same. $MU might be having its own “NVDA moment.” Red days aren’t risk — they’re inventory.

@saxena_puru Micron-Google is in its NVIDIA-DeepSeek moment. I love this dip!

$SNDK $MU Memory stocks are like sitting ducks! All parabolas end the same way...with a painful bust. Note the gap between the stock and the 40-wk ma.

@athuinvests @StockSavvyShay me too😆

@StockSavvyShay Micron-Google is in its NVIDIA-DeepSeek moment. Love this dip! Do red lights make you happier than green lights sometimes? I feel like a child going to the shop buying candy.

$MU now trading below 4x earnings 😳

@StockSavvyShay Well... look at it like NVDA with deepseek. Market reacted, and created an opportunity.

@ReyezAriel mark

google今天这篇2025年4月的论文居然引发了存储的下跌，那我们就再重读一下： KV cache 一直是大模型推理里的最大内存消耗来源。论文的做法，本质是用信息论最优的方式去压缩这些数据。不是简单地降低精度，而是重新分配信息密度。普通部分用极低比特表示，异常值单独保留更高精度。同时不再逐元素处理，而是以向量为单位编码，因为 attention 本身就是内积结构。 关键的是，它的误差已经贴近信息论下界（香农极限），也就是说压缩效率已经非常接近理论极限。论文里给出的结果，大致是 4 到 4.5 倍的压缩，性能几乎没有明显损失。效果很明显，但后续再压缩而不损伤性能的可能性已经很小。 基于大科技的内部研发流程，论文的方法及可能对模型产生的优化效果很可能已经被工程分阶段吃掉了。 比方说，低比特量化早就被用起来了，从 int8 到 int4，再到更低精度，主流模型在推理侧基本都在用。异常值单独处理这件事也不是新东西，SmoothQuant、AWQ 这些方法本质上都在做类似的事情。KV cache 本身的压缩、滑窗、分层缓存，在大模型里也已经是常规配置。 真正还没完全落地的，是论文里更极致的那一部分，比如向量量化，以及更接近信息论极限的编码方式。这些方法的问题不是原理，而是工程实现，GPU 不友好，延迟控制难，稳定性和泛化也更复杂，所以可能需要更长时间实现。 如果一定要拍脑袋猜一下论文已经落地和还没落地的部分可能有多少的话，大致可能是这么个情况：最早的 KV cache 是 1 倍成本，简单量化之后可以做到 2 到 3 倍压缩，加上异常值处理可以到 3 到 4 倍，论文再往前推一点，大约到 4 到 4.5 倍。也就是说，大部分红利已经被拿走了，剩下的提升空间不大，而且代价越来越高。 这背后的原因也很清楚。前期压缩是在去掉冗余信息，后面面对的是有效信息，再压就会直接影响模型能力。误差不再是平滑变化，而是到某个点之后快速恶化。实现难度也不是线性增长，而是明显抬升。 从模型表现可以反推，现在的主流模型已经在用这些技术。长上下文能力、推理成本下降、性能稳定，这些现象本身就说明 KV cache 的效率已经被大幅优化。像 Google 这种级别的团队，大概率已经实现了低比特量化、异常值处理和一部分 KV 压缩。 也就是说，如果说google的这篇论文对存储可能有影响的话，其大部分的影响已经被体现了出来，还没体现出来的部分，其实施难度也会较之前更大。 更重要的是，这篇论文的意义不在于多省了多少内存，而在于给出了一个边界。KV cache 压缩这条路已经接近极限，剩下的提升空间很有限。接下来真正能带来变化的，不太可能再来自压缩本身，而是需要找到其他的路径。

@billtheinvestor It doesn't make models smaller, it just lets KV cache retain more accuracy at lower quants, but KV cache is a relatively small part of memory footprint of LLMs, and model providers are already using quantized KV caching.

Google 刚刚发布了 TurboQuant —— 一个革命性的新算法。 它的作用是：在几乎不损失质量的前提下，让大语言模型（LLM）变得更小、更快。 这意味着什么？ 即使是只有 16GB 内存的 Mac Mini，现在也能在本地完整运行那些极其强大的 AI 模型了！ 而且是完全离线、免费、安全地运行。 这项技术还将带来以下重大改变： 上下文窗口大幅扩大：以前长上下文会严重变慢和掉质量，现在 slowdown 和退化大幅减少； 手机也能跑高质量 AI：未来高端手机将可以本地运行顶级 AI 模型； 速度和质量同时提升，价格却会下降。 那些以前嘲笑你买 Mac Mini 的人，现在脸都要肿了。 Google 这次把 TurboQuant 完全开源，没有自己独占。这一点尤其值得称赞——很多大 AI 实验室很可能都会把这种技术藏起来当核心竞争力，而 Google 选择把它公开，推动全人类 AI 技术共同进步。

Brent just dropped 6% this week. No ceasefire. No deal. Just expectations. Polymarket odds: • 13% US–Iran ceasefire by March 31 • 64% by June 30 The market isn’t waiting for peace — it’s already trading the timeline. 👇 What’s your call - AI matches you with a sim strategy based on your view botholder.com #Polymarket #OilPrice #Iran #MacroTrading

"In 20 major post-World War II military interventions and hostilities that we evaluated, the S&P 500 fell six percent, on average, from the initial market impact to the trough level. In 19 of the 20 events, the market took an average of only 28 days to return to where it had been prior to those events." So let’s see how this week plays out. rbcwealthmanagement.com/en-us/insights…

RBC Wealth Management looked at 20 major post-WWII military conflicts. Key takeaways: 📉 S&P 500 dips ~6% on average, but bounces back in ~28 days. ⚡ Only events hitting oil supply caused bigger drops. 📊 Short-term volatility ≠ long-term bear market. rbcwealthmanagement.com/en-us/insights… #Markets #Investing #Geopolitics #Gold #Oil

Ask for help : Has the situation between the United States and Iran really eased? #Breaking #WorldNews #GlobalPolitics

Ask for help : Has the situation between the United States and Iran really eased?

@ianbremmer The ~3,000-contract minute here was 6-10x normal for pre-news lulls😇

6:49am: sudden spike in oil futures trading. no news. no announcement. nothing public. 7:05am: trump announces a pause on iran strikes. markets move. someone knew. 16 minutes early. $580 million in contracts. the corruption is staggering.

and rising real rates