life is beautiful. to be valuable.
Metaverse
Joined January 2022
- Tweets740
- Following4.1K
- Followers110
- Likes967
Google发布第8代TPU：TPU 8tx和TPU 8i Max，其中8i Max双芯片实现计算/SRAM/HBM带宽翻倍，针对内存墙与大上下文（massive context windows and the highest compute density）；8tx针对机密推理（secure and confidential inference）。Google自家TPU在密度与安全上持续推进，将是是Nvidian强劲对手。
Today, we're expanding our eighth-generation TPU portfolio with two new systems built specifically for the agentic era — bringing our total to four distinct eighth-generation TPU systems launched in a single year.
TPU 8i Max was built for massive context windows and the highest compute density per rack, and TPU 8tx is optimized for secure and confidential inference. While flexible across workloads, the four systems in our 8th gen TPU portfolio give developers and organizations a purpose-built accelerator for every stage of the agentic lifecycle.
@AminVahdat, SVP and Chief Technologist, AI and Infrastructure, explains why we're expanding our TPU portfolio and what everyone needs to know about TPU 8tx and 8i Max ⬇️
16
vic retweeted
Andrej Karpathy:
"The biggest mistake in AI right now: people are forcing agents to work instead of mastering the harness engineering for model first.
We made that mistake in 2016 at OpenAI and it cost us 5 years "
vic retweeted
I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard.
Which harness should you run your open model in? I ran 28,849 trials to find out.
The leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness changes: the model, provider and reasoning effort stay fixed, and every task gets three attempts.
The trials ran GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint on Terminal-Bench 2.1 and DeepSWE, for $4,746 of inference.
Averaged over the three models, mini-swe-agent (@KLieret), a bash-only agent, solves 51.4% of tasks and @openclaw 27.3%. Four others are statistically tied with mini-swe-agent.
One of them is Pi (@pidotdev), at 49.4% for $0.29 per solved task, about 45% less than mini-swe-agent's $0.53.
The best harness depends on the model: mini-swe-agent on GLM-5.3-Flash, Hermes Agent (@NousResearch) on DeepSeek-V4.1-Flash and Kimi Code (@Kimi_Moonshot) on Flint.
@opencode costs $3.35 per solved task, 11.7 times as much as Pi.
See Harness Hub's Harness Lens for what the nine harnesses do differently. It documents the prompts, tools and compaction of 34 coding-agent harnesses and links every claim to its source.
Leaderboard: huggingface.co/spaces/joelni…
Harness Lens: huggingface.co/spaces/harnes…
vic retweeted
DeepSeek from @opencode is really like drugs. The speed, and how cheap it is ....
I mean sometimes I start thinking about an idea, and stop myself midway and open the terminal to start writing it out there because why waste the idea when it can turn into life for less than $1 of tokens in 5 minute
It is almost a mental disease at this point to have access to this.
vic retweeted
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
vic retweeted
I’m in love with this sentence:
“One of the most underrated skills you can learn is the ability to ignore your mood and stick with the plan.”
vic retweeted
Some news: I’m joining OpenAI to lead product management for Core Products!
I have a lot of respect for OpenAI for showing the world that scaling works, for bringing AI to over a billion people, and for believing great products are core to the mission.
I’m excited to work with @thsottiaux, an incredible PM team, and everyone across OpenAI. Together, we’ll build agents that have your back and your team’s, and keep you in control, so you get more time for the people and work you love.
Lots to learn, and lots to build. Please keep your feedback coming!
vic retweeted
Superwhisper is shipping as the default dictation setup in Omarchy 4.5. The free 10k word cloud option means you can get started without even having to download a model. Then you can switch to local-model dictation later if you like. It's so good!
vic retweeted
O mundo do software tá convergindo pra Rust.
Quando ninguém lê mais o código, a sintaxe não importa e o que conta é o resultado final.
Ouso dizer que DX (developer experience) não tem mais tanta importância e o que pega agora é ADX (Agentic Development Experience).
Arquiteturas de "software" (ou sistemas) precisam se tornar amigáveis ao desenvolvimento via agentes.
Isso (geralmente) significa:
- Self hosted runners (pra fugir do limite e custo agressivo dos runners nativos do GitHub)
- Validações automáticas (+ CI, storage e compute)
- Monitoramento
- Feature Flags
- Estratégias de deploy/rollback eficientes
Code fast, act faster
vic retweeted
Meta builds humanoid robot with Chinese startup Fourier, The Information says reut.rs/4ecTdip
vic retweeted
no matter what agent you're using, your voice is the best way to talk to them
and hex is the best way to use your voice:
- fast
- runs locally
- works on mac, windows, linux
stable release now available: anoma.ly/hex
vic retweeted
Notebook has just unlocked over 700 textbooks from the world’s leading publishers. Every single one is available to you, completely free of charge! @Gemini_Notebook
Tech or product is just part of the game, and the market or customer is another part.
programmers tend to be weak at product because they become enamored with their decisions
for 99% of decisions all of the options are valid. they just determine the position your product takes
and the decisions you can brag about usually put your product in weak positions
25
Crazy
🚨 CONGRESS IS NOW DISCUSSING WHAT HAPPENS IF MAJOR AI COMPANIES GO BANKRUPT.
House Democrats have introduced a resolution calling for AI companies to receive ZERO taxpayer funded bailouts if they collapse.
AI spending by Amazon, Microsoft, Alphabet, Oracle, Meta and SpaceX is expected to top $7 trillion from 2025 to 2030, per S&P.
Big Tech is expected to borrow a record $420 billion in 2027 alone.
Now, lawmakers are discussing what should happen if those massive investments go wrong.
The resolution calls for:
• No taxpayer funded bailouts for failing AI companies.
• No emergency loans, cash injections or debt guarantees from the Fed, Treasury or FDIC.
• Failed AI companies to go through normal bankruptcy proceedings.
• Executives responsible for reckless or illegal conduct to face accountability.
• Government support to protect workers and retirement savings instead of corporate executives.
The resolution is not legally binding, and it does not mean lawmakers know an AI crash is coming.
But it shows the risk is now being taken seriously in Congress.
And there is a reason to pay attention.
AI companies are spending enormous amounts of money today based on expectations of future demand and profits.
If those expectations fail, the damage could spread beyond AI companies to banks, lenders and investors financing the expansion.
8