Pinned Tweet
Deepseek the value destroyer
... for hyperscalers wanting to see ROI, Deepseek has reduced the numerator by 95+%, greatly reducing the odds of positive returns
Open-source / open-weight also breaks monopoly pricing power
But for consumers?
The value has shifted to consumers
Let me say the quiet part out loud:
AI model building is a money trap.
What you are seeing here is a feature of modern AI models: there is no bounding law like Moore’s law where advances can be predictably expected. Nor has any model builder adequately publicly explained what to expect sans jargon and BS.
What this means is that it’s very difficult to plan spending programs to create advances and will likely render entire months and possibly years of energy, hardware, R&D, OpEx and CapEx wasted and useless as new innovations are discovered.
Open Source is the clear winner.
Closed source will be forced to keep their best models secret and sell to enterprises OR try and create some incredible consumer app with it.
Having done this before, my lived experience is that this is easier said than done.
Net/net consumers will win, many venture funded efforts will not.
Guess $GOOG folks had similar thoughts
88
$ANT +17% today...
How strong is Gestalt overall?
A New stealth model is MOGGING Fable 5.2 😭
Gestalt is sitting at 1931 Elo, beating Fable 5.2 (1717) and even surpasses Human Golden (1739)
That's a 218 point gap over Fable 5.2 on frontier design
and It's not anthropic, xAI, Meta, openAI, SSI or Google
who is behind gestalt???
108
No need to be too hard on the guys here - after this cycle, they'll come out to be better analysts on the other side w/ both bull & bear mkt experience
Replying to @GitBizzyBabay
He's really very knowledgeable and was early to the memory trade. Unfortunately, his skillset doesn't translate to a bear market very well.
1
1
88
Alpha Exponent retweeted
Wie ich schon Mehrfach gesagt habe, glaube ich nicht an das Geschäftsmodell von Anthropic und OpenAI. Revolut hat mit 2000 Softwareentwickler auf eine Chinesische KI/LLM gewechselt (GLM) und betreibt jetzt 56 Hochleistungsgrafikkarten und haben damit die KI Kosten auf 1/10 gesenkt. Und diesen Schritt werden wir in naher Zukunft regelmäßig sehen. Die Alternative ist, Anthropic / OpenAI senkt die Kosten massiv, aber dann werden deren schwindelerregende Verluste noch höher.
Even a quantized 8-bit 27B local model can beat Opus 4.8?
Would most tasks (or most folks) require a model stronger than Opus 4.8?
What times we live in
90
If you're hiring for the avg office role
You'll not pay $300k as you might for an ML PhD
Your budget might be $60k
And you'll look for someone w/ their wits about them, but also reliable & WITHIN budget
Human work is like this, so work for AI will also likely be like this
たった今、世界のAI覇権のバランスは完全にぶっ壊れました。これほんまにやばいです。
週明けのAI関連株は大暴落どころの騒ぎじゃないかもな…。
何が衝撃的なのか、ガチ解説します。
もう何もかもが変わっちまったんだな...。
■何がおきたか
DeepSeek v4.1-Flash 1017というモデルが公開されました。世界のAI情勢を破壊します。ヤバいモデルです。
このモデル簡単がどういうものなのか説明すると、
フロンティアモデルを一部で越える(SOTA級)のモデルを中華ラボがいきなり出してきた。
問題はその性能、値段とサイズと仕組み
どれも本当に破壊的なものだから、1個ずついきます。
■性能
全分野でOpus5.5を超えたっていうわけじゃないけど。
完全にDeepSeek v4.1-Flash 1017はSOTA級
DeepSWEっていうベンチが一番衝撃的かな。
- 今回のFlash:成功率77%、約0.20ドル/タスク
- GPT-6.1 Solの最も高得点の設定:約75%、約0.65ドル/タスク
つまり、この比較では約3分の1の費用で、成功率も少し高い。大量に実装・修正を回す用途には大きな差。
というか、成功率でSolを越えてるっていう事実がヤバいんだよ。
■値段
もう完全に乱暴に言うと。ほぼタダです。<言い過ぎ
INPUT$0.15Output$0.3
入力10万+出力1万トークンの例だと
Solより約86〜93%安く、Opusより約93〜97%安い
DeepSeekハーネスを使えば、自分の利用では、入力のキャッシュヒット率が99%になることもあったので。
実際はもっと安い。
過去に湯水のように使えるAstra,Opus,Fableをくれとは言ったが。ここまでやれとは言ってない。
本当に、なんだろう、マジで安いんだよ。
実際の感覚だと200倍ぐらい安いね。
■サイズ
これだけヤバいモデルのオープンウェイトが同時に無料公開されました。
そのサイズは763B。重みだけで約510.3GB
例えば米国のフロンティアモデルのサイズは10Tとか言われてるんですね、公開情報じゃないからわからん無いけど。Kimi K3とかが2.8T
要するに、めちゃくちゃサイズが小さいんですよ。
このパラメータ数って多ければ多いほど計算に計算資源と電力がかかるんで、コストかかるんですよ。
アメリカのAIモデルって、バカみたいに計算資源を保有してるからゴリ押しでそれを解決してるんですけど。
こんだけ軽かったら、もう普通に家で使えるレベルなんですよ。
実際、DGXSpark × 4台とかMacStudio 512GBで動かせると思います。
バカみたいなデータセンターがなくても、フロンティアモデルが動くようになっちゃったってわけ。
■仕組み
ここがかなり驚き度の高いポイントなんですけど。
DeepSeek-V4.1-Flash自体はもともと存在していて。
まあそこそこ人気だったんですけど。本当に安くてそこそこ性能いいから、結構人気あるよね、ぐらいの賢さだったんですよ。
そのもともとあったモデルを自己フィードバックを使うRLで強化した。って公式は言ってて。
まあ詳しい説明は省きますけど。要するに、既に存在していたモデルを後から調整していくようなやり方で、とんでもなく性能をあげた。
驚くのは、既存モデルの後学習で、開発タスクの成功率を上げながら、処理時間や出力量まで減らしたところです。性能を上げたうえで、使う側の費用まで下げてきた。その性能の上り方が半端じゃないんだよ…意味わかんねぇよ…
DeepSeekのラボが今世界で最もAI、優れたAIラボであるということを認めなきゃいけない。
1
325
Monetization matters
AI chatbot platform Character. ai filed for Chapter 11 bankruptcy on Monday and announced it had received a $20 million opening offer to license some of its intellectual property from Disney. However, Disney previously accused the the AI company of unauthorized use of its copyrighted characters.
Read more: forbes.com/sites/siladityara…
📸: Algi Febri Sugita/SOPA Images/LightRocket via Getty Images
83
Alpha Exponent retweeted
CoreWeave’s unsecured notes are officially 1,000bps STW, now yielding ~15%.
Either this is terribly mispriced and a buy or the HY market is telling you something about the AI build-out
Which one is it?
robinhood:0x5f10a1c971b69e47e059e1dc91901b59b3fb49c3
iran strike options sound like a fancy class of exotics where strike K ∈ {K_high, K_low} is not known until the event happens & can still change ...
*WHITE HOUSE SEEKS IRAN STRIKE OPTIONS BEFORE MIDTERMS: ATLANTIC
*TRUMP EYES MILITARY ACTION TO PROJECT RESOLVE, CURB GAS PRICES PRE ELECTION
*NO FINAL DECISION MADE ON SCALE OR TARGETS OF POTENTIAL IRAN STRIKE
*LARGER MILITARY CAMPAIGN AGAINST IRAN WEIGHED POST-MIDTERMS
53
It finally happened
America has entered the frontier open weights model competition. We aren't rolling over to China
Reflection announced Beam, a 501B parameter open weights model that is comparable to GLM 5.2, Qwen 3.8, and Opus 4.8
Yes, Opus 4.8 on your desk
You will need a decent sized machine for this. Probably the Mac Studio 512gb
But if you listened to me back in January when I warned you days like today were going to come, then you're all set
I stand by Opus 4.5 being the most important model release in history. We will now have a model BETTER than that on our desks
If you haven't gotten into local AI yet, it's officially time to do it
Take any computer you have, talk to an AI agent, ask which models you can run
Never been more important to just start
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam: reflection.ai/beam
There's strong temptation for incumbents to benchmaxx - all companies suffer from this, but given the additional internal competition, teams w/in incumbents even more so...
$GOOG
I think it will be disappointing in practice but I don't care super much, what matters is that Google finally had the balls to release a completely new pretrain. Now they can do 4.1, 4.2 etc for a while and catch up like Meta did.
Other curses of Google likely remain though
1
121
There are two ways AI progress could go very badly and that we must avoid.
First, we could lose control of the future to AI. This is unacceptable; we are unapologetically on Team Humanity, and AI must always serve people. To ensure that, we need ways to ensure that alignment and safety techniques stay ahead of progress in model capabilities.
Second, we could end up in a world with too much concentration of power. If an extraordinarily powerful AI is used by one person or company to impress their worldview onto everyone else, the results could be extremely dystopian.
Avoiding these two threats requires walking a narrow middle path; for example, one country could gain too much power. Another example is one lab ending up with too much power.
84






























