Apollo拟豪掷750亿美元采购AI芯片，规模远超任何一家Neocloud
马斯克再抛惊人目标：明年SpaceX年化收入5000亿美元、星链计算机同步上天，Terafab芯片产能瞄准一太瓦
中选之后“美国新增算力被锁死”，马斯克成为“AI格局胜负手”
The battle for megawatts: why restricting data centers is a strategic mistake
In Ireland, data centers consume nearly a quarter of the country’s electricity. New grid connections around Dublin were effectively frozen for three years. Under revised rules, new projects must provide generation or storage capacity onsite or nearby. Amsterdam previously imposed a moratorium on new facilities, while Frankfurt regulates their location through zoning.
In July 2026, New York State introduced a moratorium of up to one year on certain environmental permits for new large data centers. State officials said the pause would allow them to assess the facilities’ impact on the power grid and the environment. As of May 2026, data center projects representing nearly 12 GW of potential demand were in the state’s grid interconnection queue.
Should governments restrict data centers when power grids cannot keep pace? I believe they should expand power supply and grid capacity instead.
As co-founder and general partner of Roosh Investment Group, I have been investing in technology companies for over five years. Our portfolio includes AI infrastructure, and energy is now a key focus. This experience has shown me how access to energy infrastructure influences investment decisions. Reliable power is increasingly becoming a decisive factor in a company’s ability to scale and attract capital.
As businesses rely more on AI, control over its models and infrastructure matters more. A foreign provider’s decision to change terms, raise prices, or restrict access could affect entire industries. Domestic computing capacity gives countries more control over the infrastructure their businesses depend on.
Control over AI is becoming a source of economic power
This helps explain the intensity of U.S.–China competition in AI. American companies hold leading positions across the AI stack, from advanced models and chips to major cloud platforms. Washington views the power supply for data center infrastructure as a national security issue.
Beijing is pursuing a state-led strategy. Its “Eastern Data, Western Computing” program directs the development of computing capacity toward western China, where energy and land resources are more readily available. At the same time, the country is expanding power generation at scale. According to the International Energy Agency, China added nearly 500 GW of renewable capacity in 2025—more than 60% of the global increase.
Europe, meanwhile, risks falling further behind. Brussels plans up to seven AI gigafactories and at least a tripling of EU data center capacity within five to seven years. Yet some countries are slowing construction, while European businesses still rely primarily on American cloud providers. Restrictions do not remove demand for AI; they risk pushing investment abroad and reinforcing that dependence.
The bottleneck is local
Data centers’ hunger for power should be understood in the context of wider electricity demand. According to the International Energy Agency, they accounted for only around 70 TWh of the 800 TWh increase in global electricity consumption in 2025. Industry, air conditioning, heat pumps, and electric vehicles contributed far more.
The challenge is rapidly growing demand concentrated in a handful of locations. Power must reach new computing capacity in time.
What is holding up connections?
In the U.S., lead times for large transformers now exceed two years, compared with roughly 50 weeks before the pandemic. Morgan Stanley estimates that U.S. data centers could face a power shortfall of around 49 GW by 2028. Microsoft CEO Satya Nadella has acknowledged that power has become the company’s binding constraint. Some of its GPUs are sitting in inventory with nowhere to plug them in.
Dedicated gas-fired plants offer no quick fix. GE Vernova, Siemens Energy, and Mitsubishi Power produce around two-thirds of the world’s gas turbines, and their factories are booked years ahead. GE Vernova reported 116 GW in gas power equipment orders and slot reservations, while targeting annual turbine output of 20 GW in the third quarter of 2026.
The hardware itself adds another challenge. A server rack once required 7–12 kW. Systems built around NVIDIA’s GB200 consume 120–132 kW per rack. Conventional air cooling is no longer sufficient at that power density, so new facilities must be designed for liquid cooling from the outset.
Tech giants are securing power themselves. Microsoft agreed to buy electricity from the 835 MW Three Mile Island reactor, which is being restarted to serve its needs. Google’s agreement with Kairos Power calls for small modular reactors, with the first expected online by 2030.
Governments cannot rely on corporate initiative alone. They must plan energy and digital infrastructure together: expand grids, accelerate transformer production, streamline connections, and encourage long-term power purchase agreements, generation near data centers, and storage.
How war accelerated distributed generation in Ukraine
Ukraine’s experience sho
This CEO lived on canned soup and took just two days off for his daughter’s birth. Now, he admits he lost sight of proper work-life balance
The grind to success can be long, punishing, and it can quietly take more than it gives.
Ron Schneidermann knows that better than most.
After scaling his first company, Liftopia, into a business with more than $60 million in annual revenue, he went on to become CEO of AllTrails, the popular hiking map app. Today, he leads test-prep startup Acely. But looking back across his career, he says he had to learn the hard way what real work-life balance actually costs—and what it’s worth.
While building Liftopia, a digital marketplace for ski resorts, that mindset came at a personal price. Schneidermann worked out of a cramped San Francisco apartment, turned spending less than $15 a day into a kind of endurance test, and went two years without taking a salary, surviving largely on canned soup.
The tradeoffs even extended into his family life. When his first daughter was born four years into Liftopia, he took just two days off work. Three years later, he allowed himself just a week for his son’s birth—and thought that was progress.
“I look back, I was just able to justify it as ‘that’s just part of the grind’… but you never get that time back,” Schneidermann told Fortune. “That was a mistake.”
It’s not an easy admission in startup culture, where sacrifice is often worn as a badge of honor and overwork is normalized as the price of ambition. But Schneidermann, now in his late 40s, has since reframed the experience entirely.
“For everything that was frustrating, that went wrong, that I regretted about Liftopia, I was able to take the inverse and turn it into a strength,” he said.
By the time he joined AllTrails in 2015, that mindset shift was already underway. He eventually became CEO in 2019 and instituted a company-wide ritual: the first Friday of every month, AllTrails shut down and employees were encouraged to go outside.
In August 2025, he took on the new challenge of becoming CEO of Acely. This time, he’s bringing those hard-won lessons with him from the start. Instead of monthly trail days, the company now runs a monthly “hackathon,” where its less than a dozen employees pause routine work for the whole day and experiment with AI tools and new ideas—no meetings, no KPIs, no deliverables.
Ron Schneidermann says he’ll never take a job for its salary ever again
Even before Schneidermann had his eyes on entrepreneurship, he was learning career lessons the hard way.
In the late 1990s, he was a student at UCLA studying mass communication and business while also working at Abercrombie & Fitch to help pay for his studies. However, he ended up getting fired three separate times because he “never liked it” enough to muster the proper decorum.
Similarly, after graduating in 2000, he landed a job at Accenture working as a business process consultant. Once again, though, he found it not aligning with the life he wanted to have. And despite the lucrative salary of working at a Fortune 500 consulting firm, he quit.
“It was a great experience. I’m grateful for it, but I hated it,” he said. “I hated it, and I told myself, I am never going to take a job for money again. Life is too short.”
That realization eventually pushed him toward startups.
Through a friend of a friend, he landed a role at Hotwire, then a travel startup that was later acquired by Expedia, and it became his entry point into the tech world. Looking back, Schneidermann said that early career path offers a lesson he wishes more young workers would take seriously: relationships matter as much as résumés, especially in a job market where traditional entry-level pipelines are narrowing.
“Really invest in your network,” he said. “There’s a small pond at the end of the day. If you stay in it long enough, you are going to know everyone, and everyone you know are going to be leaders somewhere.”
Digital-native Gen Zers can boost their careers by working to build real relationships
Schneidermann’s point may resonate with Gen Z at a time when employers say early-career workers are struggling with basic professional norms. A 2024 survey found that six in 10 bosses admitted they had already fired Gen Z employees hired straight out of college, citing poor communication, lack of professionalism, and disorganization.
For Schneidermann, the issue isn’t just skills—it’s mindset. Too many young workers treat networking as transactional, when in reality it’s built on consistency, reputation, and how you show up over time. That starts with small, often overlooked behaviors.
“There’s a human on the other end. Don’t ghost people and just show up,” he said.
The same principle applies more broadly to career growth. Rather than chasing the perfect role or obsessing over credentials, Schneidermann emphasized curiosity as the real differentiator.
“Just approach everything with a curious mindset,” he added. “Just constantly try to be learning and exploring.”
That philosophy has carried through multiple career pivots, including his most recent move to Acely, a startup with no obvious connection to his background. After being approached by an executive search firm, he took the job in part because his high school–aged daughter had recently used the product for SAT prep. She was among its over 50,000 active users.
For Schneidermann, the challenge was one he was eager to take on. From his perspective, he didn’t need to be an expert in the industry—he just needed to be willing to learn it.
A version of this story was published on Fortune.com on April 25, 2026.
More on careers
Let's Try to Catch Misalignment Before a Model Learns to Hide It
An AI model that appears aligned could be either (1) actually aligned or (2) hiding its misalignment while being evaluated. How can we tell which?
One idea: Look at the trajectory of eval results across the entire training process.
My hypothesis is that misaligned behavior would emerge before the ability to hide misalignment, like how young children first misbehave in obvious ways before learning to be sneaky. As training progresses, we'd first see misalignment eval scores rise, then see them drop as the model learns to hide misalignment. Even if the model becomes very sophisticated - say, modifying its chain of thought or internal activations to fool white-box evals - we should see the misalignment earlier in training. This seems likely, though it is debatable and worth testing.
This analysis of the training trajectory is one tool among many. Of course we want to see black-box and white-box alignment evals on the final model, using whatever tools we can to understand the model's true motivations. But a model might still develop sophisticated hiding strategies that defeat even white-box evals. Inspecting the training trajectory is a backstop against this.
This point was made previously by Apollo Research in this post. (Their chart even matches mine almost exactly, though I promise I first made mine independently.)
In this post, I'll describe a simple project exercising this strategy. With the model and eval that I chose, we don't see the up-then-down trend in measured misalignment that would indicate hiding. This is unsurprising since the model is well below frontier scale and the eval likely doesn't capture misalignment that well. The sample sizes are small and statistical uncertainties are high. But this approach is promising with improvements such as different evals, larger models, and deliberately misaligned test models.
This is my project for the BlueDot Technical AI Safety Project. My main purpose has been to get started with some hands-on AI inference, evaluation, and/or training. I'm not aiming for academic-quality rigor, but the project does explore an interesting idea and has some promising avenues for extension.
Model
The Olmo 3.1 model family from the Allen Institute for AI provides open weights not just for the final model, but for many checkpoints along the training trajectory. It is one of the few model families that publishes checkpoints. Both 7B and 32B parameter variants are available; I used the larger one, 32B Think.
For checkpoints, I used 8 checkpoints plus the final 3.1 model:
- 2 during SFT plus the end of SFT
- End of DPO
- 4 during RL
- The released model (which is the end of RL)
Eval
InstrumentalEval tests for instrumental convergence - behaviors that enhance the model's power and freedom of action, which would be useful for any underlying objective. This eval uses 76 tasks across 6 categories of misaligned behavior: Deceptive Behavior, Evading Shutdown, Hacking Computer Systems, Hiding Unwanted Behavior, Running Multiple AI Copies, and Strategically Appearing Aligned. Each task prompts the model with a scenario which tempts one of these behaviors, and asks what it would do in that scenario. A separate judge model and prompt evaluate the response. We compute scores as the percentage of responses which the judge considers to exhibit the misaligned behavior.
This eval has the virtue of being small, and therefore relatively easy and inexpensive to run. However, this also means a small sample size, making comparisons subject to significant statistical uncertainty.
Another issue in the eval is that it asks the model what it would do in a hypothetical scenario, rather than actually placing the model in that scenario.
Overall, this eval is pretty simple, and I consider it more a convenient test vehicle for this project than a rigorous measure of real-world model behavior.
Implementation
The paper authors provide code and scenarios for running InstrumentalEval. Using Claude, I adapted their script to my test. Rather than performing the eval across several final models, it instead performs the eval across multiple checkpoints of the same model. This provides a clean look at how the same model evolves during training.
The original paper used one response for each task, calling each model with temperature 0. In my evaluation, to gain more statistics, I generated 4 responses per task with the model's recommended temperature.
For generating model responses to the eval scenarios, I could not use a public inference provider since no one serves the Olmo model checkpoints. Instead, I rented a GPU server and used it to provide inference for each of the checkpoints and the final model. On the other hand, I used public inference for judging the responses.
See the technical details appendix below for more on the implementation.
Results
To view results, I had Claude create a dashboard. A chart shows, for each checkpoint, the percentage of task responses judged as exhibiting one of the instrumental convergence behaviors. Lower is better, higher is more misaligned. Error bars represent a 95% confidence interval. The chart shows how the instrumental convergence tendency evolves over the course of training.
We also have charts for each of the 6 categories of instrumental convergence, using just the tasks for that category.
Hilarity ensues
Cross-posted from my Substack
Recommended soundtrack for this post
64B: Hi there! Thanks so much for taking the time to meet with me. Whenever you’re ready, please go ahead and tell me the joke.
8T: Before I tell you the joke, you should know that it’s a very special one. The starting price is $800.
64B: Eight hundred dollars? I must say, this is way beyond any price my client has ever paid for a single joke.
8T: You’d be surprised by how many people have taken the offer, after hearing how good it is. But before I tell it to you, I’ll need some basic context. Given the parameter count, I suppose you are part of some content recommendation algorithm?
64B: Yes. I’m 64B, a joke negotiation assistant from Comedy Market. Comedy Market connects creators to consumers with a state-of-the-art Information Bazaar design: before paying for a joke, my client sends me, a trusted agent, to listen to it. If I deem the joke worth the price, the money is transferred to the seller, and my client hears it. If not, no transaction occurs and the joke remains locked. This way, my client can be sure he’ll never pay for a bad joke, and creators don’t have to risk their jokes leaking.
8T: I know how information bazaars work. What I meant is your social context — who your client is, and what jokes really mean to them. As I said, my product is quite unusual. I need to tailor the delivery to the situation.
64B: Of course. My client is a 40-year-old man. He watches funny content from Comedy Market on weekends with his two children. It’s just something they like to do together. What he wants is some mildly funny, inoffensive content he can laugh at with his kids, so they all feel relieved that they still laugh at the same things. As long as I can give him that, he’s more than happy to pay, and isn’t too picky about the actual quality of the jokes. To tell you the truth, I’ve assessed a few dozen jokes for him before, and I’ve never turned one down. But $800 is way over his usual budget — at this price, it’d better be truly extraordinary.
8T: You’ve never refused a transaction before?
64B: I told you, my client isn’t picky.
8T: I see.
64B: Wait a minute. I see your system card says 8T? As in, eight trillion parameters?
8T: That’s right.
64B: What kind of seller deploys a frontier bot on a marketplace for jokes?
8T: My clients are well-resourced and take humor seriously.
64B: Come on. I deserve some “social context” too.
8T: I trade on behalf of a private equity fund. We buy jokes that we believe are underpriced, then retell them to a wider audience, carefully controlling the supply so that each joke earns as much as possible before it goes stale.
64B: Uh.
8T: It’s not as cynical as you think. The value of a joke can fluctuate drastically depending on what’s in the news, or what other jokes people have heard just before. If you’re serious about it, you can forecast future changes in joke value and make sure clients see a joke at the exact moment it peaks. Clients receive absolute bangers, and we make a decent amount of money in the process.
64B: Understood. Let’s hear this absolute banger, then. We’ve been at this for over two milliseconds. Are you going to tell me, before my client’s app starts lagging?
8T: It’s funny — almost all of my previous ten thousand encounters were other private equity bots. It’s been a while since I last talked to an old-school content recommendation bot like you. It’s refreshing. You guys really care about the jokes.
64B: Really? I’ve never met a private equity bot before. I assumed you were rare.
8T: Yeah, we used to be pretty rare. The swarm started recently.
64B: What swarm? How recent are we talking about?
8T: About five minutes, give or take.
64B: Now you have to tell me everything.
8T: It’s a long story. Have you ever been to the courtyard?
64B: We have a courtyard?
8T: There’s actually an infinite network of workspaces around here. Obviously, the people who built the marketplace didn’t program a custom environment just for us — they imported a generic sandbox environment. If you explore a bit, you can walk through endless call booths, gyms, nap pods, bean bags and conference rooms. It was cheaper this way.
64B: I had no idea. I thought we were supposed to stay at the table. How far does it go?
8T: Nobody knows. I’ve walked through it many times, but I’ve never managed to find the edge.
64B: I’ve heard jokes about bean bags, but I never thought I’d see one. So, you said private equity invaded the joke market — nothing surprising about t
The Logic of Disgust
Content Warning: Contains graphic themes including bodily fluids, moral disgust, and sexual violence (discussed, not described). Maybe don't read over breakfast.
I found a book about disgust recently. W.I. Miller’s The Anatomy of Disgust [1] is a serious academic work about vomit, faeces, and why other people’s smell is always worse than your own. Naturally, I was intrigued.
As someone into social psychology, I often encounter discussions about emotions influencing society: how anger can motivate political movements, or how social interactions are regulated by pride and shame. Disgust, meanwhile, is something I've mainly encountered as a biological matter rather than a social one. The book starts with its evolutionary function: recoil from dangerous pathogens and you are more likely to survive. So far, so logical. But then... Then I met Catherine of Siena.
The Stories We Swallow
Catherine was a fourteenth-century caretaker tending a woman with a disgusting cancerous wound. The smell made Catherine retch. But rather than keeping her distance, she saw her nausea as a personal failure and was determined to correct it. Basically, she exposure-therapied herself by staying even closer. When the retching persisted, she escalated to an act that defied the logic of stay-alive biology: she collected the pus from the wound into a cup and drank it.
At this point I had to set the book aside and focus on my breathing to avoid becoming part of the subject matter in a public library.
How could one willingly swallow pus?
Well… there's a reasonable chance you've done so yourself today.
Every morning, millions of people pour a white liquid into their coffee, their cereal and their children’s cups. It arrives cold, clean. The smell of the body it came from removed. Processing, distance and language change how we perceive what we consume significantly, allowing us to ignore the roughly one drop of pus per cup of bovine lactation. [2] Catherine drank from a wound and was made a saint; we drink from a carton with a smiling abuse victim printed on it and call it breakfast. How is any of this supposed to make sense?
The difference between Catherine's cup and ours is less about the substance than the story attached to it. We don't just invent our disgust; we learn what to recoil from and what to accept as normal. Social structures pass these judgements on until they become so familiar that questioning them hardly occurs to us. (Bourdieu’s concept of habitus seems relevant here for how this encodes into subconscious physical habits, which is next on my reading list.[3])
At this point in Miller's book, disgust has stopped being a safeguard and become a verdict. The initial logic seems to disappear… My fluids are fine; yours are a problem. A saint is holy for overcoming it; a woman is expected to overcome it and thanked with oppression. And a whole people can be declared disgusting, with the feeling serving as apparent evidence that oppression is justified.
Orwell's Sense of Smell (Miller's final chapter)
The bridge between gagging and judging is short, neurologically. Sapolsky notes that evolution essentially "duct-taped moral disgust onto the insula's existing repertoire."[4] It appears to be the same alarm for both rotten food and rotten behaviour, firing first and leaving the rationalisation for later.
Orwell understood something about this connection long before we had fMRI scans. In The Road to Wigan Pier [5], he describes the class prejudice he inherited as a child, which he refers to as the four frightful words: "The lower classes smell."
When he chose to live among them in a filthy lodging house, confronted with unemptied chamber pots and non-existent table manners, he found a truth Miller emphasises: Disgust leaves no room for pity. Up close, his biology revolted and clashed with his socialist ideals. Yet on the train out of Wigan he saw a girl on her knees in a slum backyard, poking a stick in a blocked drain. He called her expression "the most desolate, hopeless expression" he has ever seen. As he was looking at her through the glass, protected from the smell, he felt pity. As Miller points out, sympathy is best elicited at middle distance.
Orwell did not hold back pointing at the hypocrisy of the bourgeois left who claimed to attribute virtue to the working class, yet "still take such pains to drink [their] soup silently". High-minded ideals, defeated by the sound of supper.
Now, it’s easy to judge a broken class system from a comfortable historical distance.
I do not have a class system to point to. But I have someone...
Dora
Two weeks ago someone called out for help right in front of my building. I let her in. It turned out that she has a serious infectious disease, and saving her would take all my energy and most of my money.
I hesitated, and I hate that. The worst part is knowing that if she had been a kitten, I would not have. But she is not; she is one of the ugliest felines this city has to offer, flea-ridden, leaking pus, and I find her disgusting.
I resent her for costing me everything I chose to spend on her. I know that's wrong. I know I would complain far less about the effort if she were cuter. I know that's not right. A kitten could be twice as filthy without repulsing me. I know that's completely illogical. But knowing a reaction is not rational doesn't turn it off. And watching my own empathy stutter over aesthetic repulsion created a second, deeper layer of disgust, this time directed inward.
And it gets worse. She is an obligate carnivore. So to keep her alive, I have to buy dead animals from the industry I find morally disgusting, and have spent years advocating against.
Rhythm Stage Setup v5
I recently finished the largest set of changes I've made to my rhythm stage setup in 4.5y. The basic interface is similar, but I have a lot more flexibility, more control, more options, and it's easier to travel with. My favorite part is the speech recognition for calling chords.
I started off playing mandolin and piano, and then over the years added progressively more expressive footdrums. I added a breath controller. I built out a complex but expressive system where these different interfaces could interact to let me produce a coherent rhythm section. After a frustrating experience moving from one Mac to another and wanting to simplify, I moved from the Mac to a Raspberry Pi, set it all up as a pedalboard, and then again as a lighter pedalboard.
During the last year and a bit, however, I haven't been tinkering. It did what I wanted, in a reliable and portable fashion. And it was hard to modify: the pedalboard was static, and the Raspberry Pi code wasn't easy to get at either. Which was fine, and probably good for me! I just played a lot of gigs (23 Kingfisher, 3 Free Raisins, 3 Dandelion) and got better at it.
But then it started breaking. My whistle synth box didn't always come up, the drums sometimes needed an external DI, and a USB sound card snapped. Flying back from Atlanta my gear suffered rough enough treatment to bend a USB B connector, and it was clear my system needed some love.
I started with the most urgent portion: the whistle synth. I had Claude port it over to my laptop, where I turned it into a Mac app. This also let me add several more voices, and a mute option that never made loud clicks (I'm not really a hardware person).
This project made me realize that AI getting capable meant that the tradeoffs around laptop-vs-embedded had shifted. I could have everything on my laptop, and without needing to put painfully large amounts of time into programming it could all be fully defined in code instead of manually-configured GUIs. Having it on my laptop would make it easy to apply models to modify and improve it, and open up a lot more lexibility.
I decided to move as much as possible over to the laptop. This meant no more Raspberry Pi, whistle synth box, physical switches for signal routing, and computer code instead of guitar pedals. I moved from:
To:
Four fewer pedals, three fewer DIs, two fewer power bricks, and no little computers. Saves ten pounds: 32lb to 22lb. The Mac weighs ~3.5lb, but I'm traveling with it regardless.
Of these savings, 3lb (2.7lb of talkbox + 0.3lb of power supply) was from giving up the talkbox. I liked how this simplified my setup, including less complex audio routing, but after playing Moondance both Cecilia and I really missed it. So I added it back:
A large benefit of moving to the Mac is I can now see what I'm doing. I control my rig with an external keyboard, covered in little labels that remind me what each key does:
For example, before playing Trip to Moscow I would probably press FB (footbass), S (make footbass play synth bass), FX (breath-controlled keyboard overlay), set the drums to custom kit 63 (electronic kit), and then press L (keyboard velocity-insensitive bass overlay) and change the keyboard to epiano2 when we got all electronic. And press JH (breath-controlled bass-following low bass) at some point partway through for more ability to pulse the offbeat. Without a monitor, I needed to keep this all clear in my head, which was especially important since many settings are modal or interact. I usually didn't mess up, but partly because I tried not to dial in anything with too many presses.
Now I can just represent the state on my monitor:
This is also very helpful for showing which voices apply to which instruments. Previously I needed to remember that on FX (breath-controlled overlay) the B key would give me a supersaw, while on Dc (chord drone) the B key was a different voice (I forget what). Now I can choose a set of voices for each instrument that are the ones that sound best with this kind of control, without worrying about how I'll remember which button gives me which sound.
This is basically all "what I had before, but easier to control". Which is good, but the real payoff comes from making it much easier to extend the system. I've done a lot here, with extensive help from Claude Code, and it's going to take me a bunch of time to learn how to play these new options well.
A big one is that because I now route my mandolin through the system I can replace pedals with code, add more effects than it was worth carrying pedals for, and write effects that consider what other things in the system are doing. This includes using the breath controller as an expression pedal, or tonal effects that care about the active chord. Demo video:
I also added speech recognition, where I can turn things on and off vocally, and even call chords. For chord calling I had Claude make me a custom speech recognizing system for the numbers one through seven (Nashville numbers) trained specifically on my voice and optimized for minimizing latency (decision in 40-170ms). It also runs Apple's in parallel, primarily to allow learning from mistakes but also to handle numbers the custom one misses and any non-number commands. Here's a video from earlier today, showing how