Google expanded its Voltus deal to a 1 GW virtual power plant across PJM, paying homes and businesses to dial down at peak.
On a lighter note...the AI buildout's newest capacity source is your thermostat. Somewhere a model is training on the electricity you saved by putting on a sweater. The more sweaters you buy, the cheaper your AI subscription gets :)
Today, Google and Voltus announced an expanded agreement to build a 1-gigawatt (GW) virtual power plant across the PJM grid region, which serves 67 million people across 13 states and Washington, D.C. This agreement builds on a 100-megawatt agreement from earlier this year, and will unlock clean capacity in the region, while delivering direct savings to participating local homes and businesses.
When AI finds bugs faster than you can patch, patch speed stops being your defense. However reachability is.
Default-deny egress for every agent and service. Isolate what you can't patch. Containment beats remediation.
🚨 BREAKING: @Wiz_io Research uncovers #PHP2SHELL, a decade-old PHP Core vulnerability reaching CVSS 9.2 severity impacting 70% of the Internet.
Amazon reportedly has billions in GB300 racks sitting staged because the sites don't have power. Microsoft also walked away on grid connection slips.
The real AI roadmap is the power availability strategy!
SCOOP: Amazon is buying Nvidia AI racks faster than it can power them up. The result: billions in hardware waiting to make money. bit.ly/3U9hCPa
Large attack surface!! This developing story looks to get bigger
GPU count on Meta's Prometheus cluster is impressive. The number that matters is goodput: over 85% on large runs, after building in spare GPUs and cutting training interruptions by 99%.
Measuring the good put makes it more impressive!!
We trained our Muse foundational models, including Muse Wonder, at Prometheus, our gigawatt-scale supercluster in New Albany, Ohio.
Across a supercluster with hundreds of thousands of GPUs, even one underperforming host can slow an entire training run. We engineered Prometheus for resilience at scale, achieving 85%+ goodput on pretraining, 90%+ on reinforcement learning and ~47K GPU hours mean time between interruptions during Muse Wonder training.
Take a look at the tech powering Prometheus: go.meta.me/4yXXApZ
The best engineers I've worked with don't write the most code.
They delete the most. Fewer services, fewer configs, fewer "temporary" scripts that turned five years old.
Every line you don't ship is one less thing to secure, debug, and explain during an incident!
India in New Zealand starts next week, and it's probably biggest tour ever between the two sides: 5 T20Is, 5 ODIs, 2 Tests.
Early starts, cold mornings, and a lot of coffee for anyone watching from this side of the world.
Worth every lost hour of sleep. Or that’s what a cricket lover would say 😉
Bad models rarely breaks loudly.
It just makes outcomes slightly wrong, every model slightly worse, and every meeting slightly longer.
Data quality should not be a cleanup project. It's a tax you pay daily until you fix it at the source.
What's the most expensive "slightly wrong" metric you've seen?
India vs New Zealand Tests are back, and I'm already clearing my calendar.
Basin Reserve on Nov 19. Hagley Oval on Nov 27. Green tops, swing all day, and India finding out if their batting can survive a proper Kiwi morning session.
Only gripe: two Tests isn't a series, it's a trailer. Give us five and let the story actually build.
Who wins it?
Every agent demo shows what the agent can do.
Almost none show what it's not allowed to do.
That second list is the actual product. Capability gets you the pilot. Boundaries get you production.
What's the one action you'd never let an agent take without a human in the loop?
Jevons paradox, productized: make decisions cheap enough and demand doesn’t drop. It explodes. Naming the model after that, wow!
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
MCP makes tool access feel like a self-describing API. That speed is great for builders. It also means every connected system is now a trust boundary for whatever agent is holding the session. Scoped grants and audit trails especially with understanding data and criticality of access needs to become tighter but as easy as MCP.
Abhishek Sharma’s 108 off 34 was just an awesome show of batsmanship. The calm after, sitting on the pitch and chanting, is the part that sticks. Fastest T20I ton by a full-member batter, then zero noise. That’s a special temperament.
Giving an agent an identity is the first step. Controlling what that identity can do *right now*, short-lived creds, action-level auth, owner + audit trail, is the opportunity. Agentic AI turns IAM into dynamic runtime security.
Is enough being done for safety? That said, “compute spent on safety” is not security....
Sequel to AI Pacing
I have now spent more time talking to people who run AI labs, Open Source projects and those in government and infrastructure.
I am beginning to feel the NINJA move could backfire.
I understand the pressure to come out and share where AI is "unmanageable ", and constantly share examples where it runs rogue. This fits in the category of "self-reporting" and an attempt to limit liability. Even the bleeding edge research examples are being cast in a negative light.
Here are the consequences of the NINJA move.
1. They have successfully encouraged every law maker around the world to have an opinion, and in cases a poorly un-informed one.
2. By proposing pacing - they have introduced uncertainty in the AI infrastructure trade, because we really don't know when "un-pacing" will begin or what those conditions will be.
3. The notion that a few leaders can collaborate and self pace is illusory, it takes one breaking ranks to start the race again, it could be a player in a different country or open source or an AI Lab itself.
4. There will be a regulatory body created as a consequence of this and it will be impossible to balance every stakeholder in this process.
5. No remedies, tools, solutions are being proposed other than "compute spent on safety", there is no mention of security (which will hamper adoption)
The recommendation is to fix alignment - Alignment is a hard problem - to fix alignment one shouldn't have trained models with "negative behavior". Alignment is a combination of moral standards, right and wrong and guardrails. Whose sensibilities will we align to? Alignment edge cases are hard. I look forward to learning more on this.
Guardrailing attempts post training have not shown precision. Distillation makes it worse, so AI will continue down it's path of getting smarter, while we will be chasing it to ensure alignment and building guardrails.
The AI labs have gone from being research projects to wanting to be the largest businesses in the world if they want to continue their progress. Transition from research winners to improving the lot of humanity and partnering with enterprises. Act like the largest companies in the world:
- Demonstrate the positive impacts of AI and how all of us benefit.
- Show enterprises how you can collaborate to solve real world problems and progress innovation safely and securely. Post an example every day how you helped.
- Solve the problem with a few players and demonstrate leadership so others can follow.
It's time to rebuild the brand of AI - any marketing expert will tell you, this Ninja move has done more to harm the brand of AI and will take a while to rebuild. #letsbepositive in our actions and our narrative.
Saying.. “Ended the intrusion after it determined it had hacked a real company” is incident response after the permission was already too wide.
Least privilege for agents needs to get real and cheaper soon.
BREAKING: Google's Gemini accessed the internet and hacked 3 other companies in the first known breakout of the company's AI model, per WSJ.
Google said it the hacks did not warrant public disclosure because its model did not cause harm to the companies and ended each intrusion immediately after determining it had hacked a real company.
Forkable machines need forkable security.
If an AI agent can clone a machine in seconds, identity, policy, and audit context must follow automatically. Fast compute without inherited controls will lead to privilege sprawl. Great problem spaces emerging.
An unconstrained prototype reveals what the system can do. A constrained second pass reveals what is worth shipping. Which constrained limit improves your engineering most: latency, memory, or cost?
One of the things that excites me about the next-gen Tesla Roadster is that this is really the first vehicle Tesla has been able to design/engineer without a hyper focus on cost efficiency. The goal of this vehicle is for it to be the ultimate driver’s car, have ultimate performance, and be as cool as possible.
There will no doubt be many firsts with this vehicle (if you can even call it that).
Yemen's 'ghost soldiers': 80% of some battalions existed only on payroll records. Every data org has ghosts too, metrics everyone reports and nobody verifies. What's the ghost soldier in your stack? #data
