Robotics Engineer 🤖💻 | Anime fan 🌟 | Gamer 🎮 | Tech innovator 🚀
Seattle
Joined September 2022
- Tweets3.1K
- Following349
- Followers23
- Likes12K
xpetri retweeted
Burning through 22.5 million tokens just to generate an asset slice for twenty cents is the ultimate DeepSeek flex, but it highlights the absolute madness of modern multi-agent pipeline optimization layers.
Dumping an entire media source file straight into an active context window looks beautiful on a pricing comparison chart, but the telemetry budget inversion is real:
• DeepSeek gives you massive token volume bandwidth for basically free, but your engineering team spends hours debugging loose reasoning frames.
• Claude Opus charges a premium enterprise license fee per million tokens, but it resolves the state machine logic parameters cleanly on the very first turn.
When building or scaling your multi-modal integration setups in your current production pipelines, do you lean toward code-first token compaction to save context window space or just spam cheaper flash provider endpoints?
xpetri retweeted
Agents should never directly see your credentials when you log in.
That's why we shipped Secrets, let agents use encrypted credentials at runtime without ever seeing them.
Functions and Secrets are free for all Browserbase users, try them out and lmk what you think!
xpetri retweeted
🎧 Full roundup: linktr.ee/odsc
🔗 YouTube: youtu.be/0J0UFGUc7og
xpetri retweeted
In Case You Missed It: Last Week in AI | October 12–18, 2026
Three US organizations released open-weight models in two days. At the same time, Anthropic reported junior technical hiring had fallen from 27% to 9%.
Also this week: Claude Code Review went GA, Gemini 4 Argon opened more broadly, and OpenAI published new misalignment reports.
#AI #OpenWeights #AIEngineering #FutureOfWork
xpetri retweeted
exciting to see more work on diffusion based LLMs
for most use cases there is little to be gained from further pushing the frontier but so much more utility and use cases open up with higher speed and lower cost
xpetri retweeted
Read more on the technical report.
skyfall.ai/blog/introducing-…
xpetri retweeted
Great post by @Leoskie_L on Nscale’s shortcomings impacting the Neocloud financing availability.
Nscale’s UK site was scheduled to be powered on in 2027, now power availability pushed back to the 2030s.
Banks have to be a lot more selective on which Neoclouds buildouts to fund.
微軟退出英國 neocloud Nscale 在 Essex 的資料中心案，表面上是一則合約新聞，其實是 AI 工廠最現實的一道關卡：GPU 買得到，機櫃做得出來，但電接不上。
照 Telegraph 和 Bloomberg 的報導，主管機關告訴 Nscale，Loughton 這塊地要到 2030 年代才有電可接，原本預計 2027 年就要上線。微軟是這個場址唯一公開的客戶，原本要在這裡租約 2.3 萬顆 NVIDIA GPU。電一延就是好幾年，這個案子等於判了死刑。但要講清楚，這不是微軟跟 Nscale 分手。英國案只佔微軟合約的一小部分，葡萄牙 Sines 和德州的合約都還在，微軟 4 月才在葡萄牙加碼 6.6 萬顆 Vera Rubin。被丟掉的是那塊沒電的地，不是這家公司。
真正讓人不安的是 Nscale 本身的結構。它正準備在紐約上市，代號 NSCL，Telegraph 說估值最高上看 350 億美元。S-1 寫合約總值約 1,034 億美元，其中微軟約 438 億、Anthropic 約 446 億，兩家合計約 85%。但到 8 月底，真正在跑的只有約 26 億。剩下那 1,000 億，要等電接上、機櫃到位、準時交付才會變成營收，挪威旗艦機房原定 30MW，目前只上線 6.4MW，進度已經落後半年。更關鍵的是，S-1 寫明 Anthropic 那批要是交付延誤，對方可以不負責任地終止。換句話說，大型雲端業者把產能外包給 neocloud，延期的風險就一起推了出去。微軟自己蓋機房，延期得自己吞；租 neocloud 的產能，延期就可以退。風險全落在 neocloud 身上，而它們的錢大多是借來的。
這也是為什麼上週整個 neocloud 族群先跌了一輪。CoreWeave 單週跌近 25%，債券殖利率破 16%，5 年期 CDS 破 1,000bp，Nebius、IREN、Applied Digital 也各跌約 9 到 10%。Nscale 上市被當成市場還肯不肯借錢給 AI 基建的試金石。這時候再看 NVIDIA 跟 Nscale 的關係就更微妙了：NVIDIA 同時是它的供應商，也是客戶（12 億美元保留算力）、投資人（10 億美元，預計 11/16 交割）和租約保證人（約 8.6 億美元）。這是典型的循環融資，Nscale 上市要是卡住，NVIDIA 這條線的曝險就會被放大檢視。Dell 也在線上，Nscale 的 GB300 NVL72 機櫃是 Dell 整櫃交付，場址延期，機櫃就得延後出貨或改送別處。
當然也可以問另一個問題：微軟真的只是因為沒電嗎？Reuters 提到，大型雲端業者的第三季財報暗示資料中心案延期、資本支出成長放緩。缺電是真的，順勢砍掉最弱的一塊產能也說得通，兩件事可以同時成立。但我不認為這代表 AI 需求轉弱。Anthropic 8 月才簽下 446 億美元的合約，需求方還在搶算力，卡住的是供給端的電和資金成本，不是 GPU 沒人要。
所以這件事最大的訊號是：沒有自備電源的場址，在 2026 年已經很難賣了。Nscale 在西維吉尼亞給 Anthropic 的園區走的就是自己發電，最高 8GW，2028 年上半年先做到 2GW。受益的是燃氣渦輪、燃料電池、變壓器和開關設備這條線，美股的 GEV、VRT、ETN，台股的台達電、華城、士電、中興電、亞力，吃的都是電網排隊這個結構性題目。但它們跟 Nscale 沒有直接合約，主要客戶在台灣和美國，不要把英國一個案子直接對到它們的營收。鴻海、緯創、廣達這些機櫃代工廠也一樣，就算有經手 Dell 這批，頂多是出貨延後或改送別的場址，訂單不會消失。
看 neocloud，別只看合約總值寫多大。要看每個場址什麼時候真的通電，有多少合約真的在跑。
xpetri retweeted
Modern #Statistics with a Computer Based Approach in Python! #BigData #Analytics #DataScience #AI #MachineLearning #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #GoLang #CloudComputing #Serverless #DataScientist #Linux #Books #Programming #Coding #100DaysofCode
geni.us/Modern-Statistics
xpetri retweeted
Not ALL job openings are legit.... Some are fake camouflaged to steal your money, time and personal information
How can you tell which is real?👇
Save for later📍🔖
xpetri retweeted
Day 19/31 🎥 Verify Suspicious Videos Before Sharing
🔎 Check who posted it first
👁️ Look for odd lip-sync, lighting or blinking
🖼️ Reverse-search key frames
⏸️ Pause before you share
💬 Have you ever shared something that turned out fake?
#CybersecurityDiet @ireteeh
xpetri retweeted
NVDA, TSM, MU: Nvidia custom HBM hits 30% more bandwidth while UBS flags 91% demand surge and 10% supply shortfall
+ Nvidia is working with Samsung and TSMC to design the base die for NVHBM, its custom HBM, per Seoul Economic Daily. Shifting the memory controller off the GPU onto the base die and shrinking the PHY 60% lifts bandwidth up to 30% and cuts power 15% versus standard HBM4E.
+ Samsung's own 4nm base die lines sit only at Hwaseong Line 17 and part of Pyeongtaek P3, too thin to absorb custom volumes. Industry estimates put TSMC 3nm yield above 90% against roughly 60% for Samsung Foundry 2nm SF2, so Nvidia is consolidating base die production at TSMC.
+ UBS projects HBM demand next year at 61.5 billion gigabits, up 91%, against shipments of only 54.6 billion gigabits, leaving a 10% shortfall. TSMC is answering with two new 3nm lines at Fab 18 in Tainan plus 3nm at its second Arizona plant and second Kumamoto plant. Samsung is hedging by validating the base die on its own 2nm process.
The demand math keeps compounding. JPMorgan puts the HBM market at $64B in 2026, $160B in 2027 and $280B in 2028, with memory shortfalls near 20% in each of the next two years. Every point of that shortfall is pricing power for whoever ships qualified HBM first, which is why base die work is consolidating at the highest yield foundry.
xpetri retweeted
OpenAI and Anthropic weekly recap, week 42 of 2026: Sora revival talks, a hidden recorder at OpenAI, Anthropic's own source control Scion, a solved ECDSA mystery and more
The Information reports OpenAI is in talks to revive Sora through a startup co-founded by Jeffrey Katzenberg and former Sora lead Bill Peebles, with OpenAI taking equity if a deal closes
OpenAI says they found a hidden recording device in an office, used their models to trace it to an employee, and the FBI made an arrest
The Information also reports early OpenAI talks to buy security firm Trail of Bits
ChatGPT now gathers information across open Chrome tabs through OpenAI's extension, and can control them in ChatGPT Work
OpenAI's GPT-6 Pro got faster with better tool use and multiple agents working in parallel
GPT-5.5 is retired from ChatGPT, ChatGPT Work and Codex, API unaffected, and OpenAI's GPT-4o and GPT-4.1 leave the API on January 14, 2027
Anthropic says strengthened defenses against distillation attacks cut what distillers can harvest by 90 to 99 percent
An internal Claude research model at Anthropic solved a 25-year-old mystery about the NIST ECDSA curve seeds, tracing them to a C library random generator seeded with Unix timestamps
Anthropic built Scion, their own source control, and moved their monorepo off git in five months as daily merges went from roughly 2,500 to over 30,000
Claude Code desktop was reworked for many sessions side by side, Code Review is generally available for Claude Team and Enterprise, and Opus 5.5 fast mode left research preview on the Claude Platform
Anthropic reports agent hours passed human hours internally in March, opened a Singapore office, and says Democracy Works voting information comes to Claude ahead of the US midterms, and more
xpetri retweeted
Anthropic opened two new fellowships. Four months at Anthropic starting January 2027. Deadline: October 28.
→ STEM Fellows: science and engineering. Help Claude reason about the hardest problems in those fields.
→ Conceptual Reasoning Fellows: philosophy, cognitive science, economics, policy, and law.
→ Both work directly on improving how Claude thinks about complex problems in their domain.
Not an internship. Not a research assistantship. Domain experts embedded inside the lab that builds the model.
Apply ↓ job-boards.greenhouse.io/ant…
xpetri retweeted
Source: opswat.com/blog/the-gap-betw…
xpetri retweeted
Anthropic just leaked the setup that turns Claude Code into a full $750k reverse engineering lab.
you drop in one file, four minutes later you have a full description of how it works, and you can rebuild your own version from scratch
Opus 5.5 runs the session and REA picks the right tool for each job: one takes the binary apart, another decompiles the code, a third watches the program live while it runs.
one command mapped 214 functions, decompiled main, pulled 1,204 symbols, traced 41 live calls and found the path in 12 steps.
then it wrote SPEC md from what it learned and ran a clean-room auto-fix against it, 42 of 42 checks pass.
the whole dock is 20 open source repos already sitting on GitHub, nothing custom, nothing paid.
save it before Anthropic pulls it from public access.
xpetri retweeted
The keys that got compromised racked up lots of tokens in a very short amount of time, this usually happens when someone uses stolen keys to generate data for distillation, not for regular black market reselling to consumers
And no distillation doesn’t imply that YOUR data is used
xpetri retweeted
If you don't receive any update from HR after your interview, send this follow up email 💌
xpetri retweeted
TOKENMAXXING: @vllm_project and @inferact achieve 1.3 BILLION tokens per $ of TCO (hyperscaler volume pricing) via disaggregated serving on AgentX.
This result is 4x better than the B300 aggregated configuration.
