nob

The world is made up of someone else's work.
So I'm interested in products, services, technology and productivity-enhancing techniques.

[nob @nobstrong]

@sep_is_heim すごいな、よく思いついた！

[shi3z @shi3z]

DeepSeek v4.1 FlashのM3 UltraとDGX Sparkでの最適化を同時並行で進めているが、最適化が佳境に入ってくる(世界最速に近づいてくる)と、急にClaude Codeがつまらないベンチマークをとりはじめ「こんなことやっても意味ないっす」「前例がないっす」とか愚痴り始めた。まるで無能な人間のプログラマーみたいだ。

こういう輩に言うことはただ一つ「うるさい。いいからやってみてから考えろ」

[nob @nobstrong]

福野礼一郎の「バンザイラン」を思い出してしまった

Nelson Racing’s 1,200 HP Twin Turbo V8 https://t.co/Y0pYBMs3hd @YouTubeより

[nob @nobstrong]

@yukihiro_matz テスト設計が要件とあっているかがわからないから

[pbt @pablopagos]

ネパール洪水、この動画が一番生々しい…

[nob @nobstrong]

このネタを聞きたい！「Qwen3.8 27Bが驚異の6.2GBに!? Unslothが新量子化技術で実現」 #backspacefm https://t.co/vLhryUJZGx

[nob @nobstrong]

@KimiAI_Japan @alfredplpl キミとなら！

[Kimi 日本公式 @KimiAI_Japan]

はじめまして、日本。
Kimiです。

今日から、日本での歩みを始めます。
これからの出会いを楽しみにしています。

抽選で 10名様に
月間Allegrettoプラン(39ドル相当)をプレゼント
リプ＋フォロー＋リポスト＋いいね

Kimi K3を体験してみたい方は、この機会にぜひ

知る、考える、つくる。
Kimiと。

[shi3z @shi3z]

Premireとは似ても似つかないが
Qwen3.8 27Bで動画編集ツールが二時間くらいでできてしまった(MacBookPro M4 128GB)。もうこれからの動画編集者は自分で動画編集ツール作った方がいいかもしれないね。好みの編集できるし。 @drikin も編集を自動化してるしな

[大曽根宏幸 / Hiroyuki Osone @OsoneHiroyuki]

@AmadeusSVX https://t.co/GqLUyW6I3w

[渡边君 @JiaweiShen2568]

比亚迪腾势Z9GT展示全球最强电控能力和无人驾驶水平
被称为特斯拉20年后的技术水平
这台车搭载了9分钟充满电的闪充技术
最后再强调一遍
这些漂移的车内没有人

[nob @nobstrong]

このネタを聞きたい！「【西川和久の不定期コラム】 Sora 2超えでSeedance 2.0に迫る！動画AI「MiniMax H3」が凄まじいクオリティだった」 #backspacefm https://t.co/vLhryUJZGx

[nob @nobstrong]

@shi3z 10年ぶりくらいに大前翁の本を買ったが、いつも通り読みやすい。
むすびにまだまだやりたいことがあるから死ねないと書いてあったのが嬉しかった

[nob @nobstrong]

@sion_mrk シフト制だと、祝日は代休としてもらえないんですか？

[下田丈 @joshimoda]

【重大発表】日本に帰ります！🇯🇵
10月10日・11日に開催される、D.I.D全日本モトクロス選手権シリーズ2026 第8戦への参戦が決定しました！🔥
日本でレースを走るのは、2024年の秋以来。成長した姿を日本のファンの皆さんの前でお見せできるのが今から本当に楽しみです！

[nob @nobstrong]

スマホのカメラ性能を高めるよりも、純粋にカメラを作ってほしいよね。

-------
1994年・浪1のとき:Apple QuickTake 100発売(民生用デジカメの先駆)

アップルブランドで売られた(製造はコダック)、家庭向けデジタルカメラの先駆け。640×480で8枚しか撮れず、画面もないので撮った写真はMacにつなぐまで見られなかったが、フィルムなし…
#SFライフライン https://t.co/6wbPEPXAP3

[nob @nobstrong]

@kenn AI人間の写像と思うようにしてから自然とやってました！

コンピュータサイエンスも必要だけどヒューマンマネジメントのナレッジをLLMに試してみるのはいいかも

[せ @Politikar_ML]

Google垢BANによる研究資料消失の件、主指導教官に話しました。
途端に泣き崩れる先生。
申し訳ありません、これ以上復元はできません。
今から副指導教官に資料消失したので学会発表が悲惨になる事、伝えます。
Googleを活用してる人もいるんです。俺はスンダー・ピチャイを絶対に許さない。

[nob @nobstrong]

1985年・小4のとき:NEC SX-2稼働(日本のスパコンが世界最速級へ)

ベクトル型スーパーコンピューターで日本勢がクレイに肩を並べ、日米の計算機競争が本格化。気象予測や航空機設計が計算で行える時代になり、この系譜が地球シミュレータ(2002)につながる。
#SFライフライン https://t.co/aku2BgIWmi

[Tech Fund @techfund1]

Nice expert call on $CBRS with a Director at Figma who's sending 40% of AI traffic to Cerebras' cloud

There are clear use cases for hyperfast inference and the customer is willing to pay a premium for these fast tokens

Figma spent roughly a year building an in-house language model for its assistant product, a chat interface intended to make Figma AI-native. The model is in the 50 to 100 billion parameter range, which he characterizes as roughly five to six times smaller than the frontier models it replaces, and better in quality for the specific task. Figma ships the weights to Cerebras and Cerebras handles all serving and optimization. He explicitly frames the relationship as a cloud provider relationship, not a hardware purchase.

The AI traffic split:

- 20% of assistant traffic routes to external frontier models (Anthropic and others) for general question answering and RAG-type requests. This share is structurally permanent. Figma has no intention of training its own model for it.
- 80% routes to the in-house model.
- Of that 80%, half goes to Cerebras, so roughly 40% of total assistant traffic.
- The remaining 40% goes to Fireworks, Modal and Baseten.

Figma Make, the vibe-coding product, sits in the same org but runs almost entirely on external models from Anthropic and Google. Critically, the AI assistant is currently exposed to only around 30% of the Figma user base.

Current Cerebras spend is described as low to mid seven figures annually. His forward view: "I would expect at least in the next one year to easily double... maybe 2X-3X growth in the next one to two years."

Beyond that he expects the use case to plateau, since the core objective is penetrating the existing user base rather than acquiring new users. This is a useful anchor for how a mid-size, non-AI-native enterprise customer scales on this platform: fast doubling off a small base, then aiming to flatten once penetration completes, with the next leg dependent on new products rather than the same product growing.

# Why Cerebras Does Not Get the Other 40% of Traffic

1. Capacity is reserved, not metered. Figma has weekday peaks and effectively zero weekend and overnight traffic. Reserving for peak means paying for idle silicon most of the week, so Figma deliberately sizes its Cerebras reservation at only 40% to 50% of average traffic and sends the peaks to on-demand GPU providers. This single design choice is what creates the opening for Fireworks, Modal and Baseten.

2. Reliability - Error rates are still above what he wants from a production system, which forces Figma to maintain fallback paths.

3. Observability - He wants first-try success rates, retry counts and internal failure data, which he is not currently getting, and without which he cannot engineer around the reliability gap.

"If Cerebras did not have any limitation, we would've used the whole 80% of traffic through Cerebras."

# The Deployment Friction, Scored 7 out of 10

Model updates are not self-serve. Figma has to notify the Cerebras team roughly a week ahead. Hand a model over Monday or Tuesday, and it is deployed by Friday. Two to three days per iteration, versus effectively immediate self-serve deployment at every GPU-based competitor.

He rated the pain at seven out of ten on a scale where ten is severely damaging.

"It's of course something we can live with, but it definitely slows our execution a lot... we build a model, we do some testing, we do this three-day wait, do some testing, find that there is a small bug, and then we have to retrain the model."

This is a developer infrastructure maturity gap, and it is the kind of thing that does not show up in benchmark comparisons but does show up in renewal conversations and in how much of a customer's roadmap a vendor can capture.

# The Price of Speed

Against the frontier model, running Figma's own smaller model on Cerebras costs roughly half to one third as much for equivalent traffic. But the model is five to six times smaller, so the like-for-like inference premium is real.

Against a GPU-based provider running the same model, his estimate is that Cerebras is around 30-50% more expensive. Figma has not run a formal side-by-side, which he attributes to the fact that reserved pricing versus per-token pricing makes the answer entirely dependent on the traffic profile.

In a bake-off against hyperscalers, Modal, Baseten, Fireworks and Groq, Cerebras came out roughly 10 to 15 times faster than most of the market. Figma found Cerebras through the Artificial Analysis public benchmark, then spent months on POCs specifically to de-risk infrastructure maturity before scaling to production.

The latency threshold he describes is a genuine product constraint rather than a nice-to-have. For assistant tasks a user could perform manually in about 30 seconds, an AI response taking longer than that adds no value. Delivering in three to five seconds changes the product.

"Speed is extremely valuable... even that 50% to 2X increase in price, I think is totally something we are willing to pay for the speed."

Cost optimization at scale would likely not come from switching hardware. It would come from shrinking the model further or routing simple requests to a smaller model.

"Maybe a workflow user makes a request, it takes three hours to run. If that's the case, then speed is not a big criteria... For those use cases, we may switch to the cheaper inference provider."

Figma has not built an in-house model for Figma Make because the output format is React and HTML, which frontier models have seen extensively, leaving little quality headroom. But he notes cost and latency wins are still available, and if cost becomes a concern as Make scales, Figma would likely train a smaller model and host it on Cerebras for the base load. Additional projects beyond the assistant are early stage, with some expected to scale in the second half of the year.

He also notes Figma experimented with Groq prior to its acquisition by NVIDIA.

$CBRS $NVDA