Derek Fulton

@AniC_dev I have been getting more than 200 with R2 by tweaking chunk size and concurrency. Also check out Backblaze less than 1/2 of R2 base rate, no egress for 3x what you have stored and I've cracked 700MiB/s there on download.

box is a forkable full-VM AI sandbox it means you can replicate any box around, whether stopped or turned on, great for making templates here you see me fork in no time a box that had minecraft installed on it with one world already in the forked box I can still open minecraft and play my world (at 10fps coz its a AI sandbox not a gaming laptop lol) snapshots are browsable and downloadable at any time via the CLI, API, SDKs and our dashboard snapshots are up to 50GB, filesystem only so > resume behaves like a server restart which is often preferred by builders, more predictable and hard to fuck up, you can use systemd like you're used to > they're kept secure in R2 in Europe, even if all our sandbox infra was to go down they'd survive > and they're just files so they're portable, readable. to restore we first download and mount via FUSE all the filesystem metadata back super fast so your fs looks and behaves identical as soon as you get access then we restore all the data in the background super fast from R2 (at ~200mb/s, there are a lot of tricks involved and footguns avoided) we're learning what restore order is best on each resume by noting what you read in the first few minutes, so the next time is always smoother

I'm convinced this is all just marketing

@mattracquet I need this data ASAP. Did you have to pay for it? Is any of it free? This is the best I can find: github.com/JeffSackmann/t… Btw, Carolina Cloud would love to host some bigger datasets publicly if possible.

Sinner was serving a bit slower than his 2026 averages (in image) today but it's funny that this matchup is often still framed as Zverev having a slight serving edge. Sinner is as potent if not *more* potent than Zverev on serve, in large part because his placement is genuinely extraordinary (tour avg is 60cm away from side and T lines, Sinner's is just 49.7cm). 44% of Sinner's 1st serves don't come back compared to 42% of Zverev's. And Sinner's deuce wide serve, and T serve on AD, especially are incredibly tight to the lines, just look at how tidy the clusters are. We've never seen this combination of serving and ground game ability in one player

@SeanODowd15 @NWischoff Very case by case. Rolling the dice. Have heard some huge success stories as well

@NWischoff Strongly recommend against. We tried it and it was an disaster

For those of you that have had or have an au pair - do you recommend it?

Claude Code with Fable is a dinosaur compared to grok-4.5 + @pidotdev. Literally an order of magnitude difference.

@the_transit_guy Take the train. Door to door is actually around 5h. Train stations are more central. No airport parking nonsense to deal with. 1/10 the price if you’re cost sensitive. Also didn’t even know this train existed despite being from Charlotte

The richest country on Earth: where a 245-mile trip between major cities means choosing between a $500 flight or a 6-hour train ride. Sure am glad we haven't invested in fast trains!

@fayazara Wow this needs to disrupt all the garbage PDF services websites. Unbelievably fast. Seriously impressive.

Burned my Fable session and compiled qpdf on wasm So you can do all those PDF actions like unlock, lock, merge, and split, right in the browser itself It also works on Cloudflare Workers btw pdfstudio-demo.fayaz.workers.dev

@BenSiranosian Local NVMe by default on Carolina Cloud. $0.01/100GiB/hr. At the cost of backups, but that’s what object storage is for.

The unreasonable effectiveness of just running your workflows on a really, really big computer with really fast local SSDs

@BraydenWilmoth So sick. Not sure why this hasn’t existed for years. Super long $NET.

Introducing Cloudflare Drop Drop your folder in the browser and deploy it instantly on Cloudflare. Your website... milliseconds away from users on region: earth No account needed. Deployment is active for 60 minutes, then expires unless you claim it. cloudflare.com/drop/

I built a voice agent to prank a friend a few years back and it was enormously painful to hook up Twilio, the OpenAI API, etc. Very nice move from @SpaceXAI.

@alexocheema What’s your over/under on months until we have a fable-tier model that can run on a sub $50k rig?

we're going to scorch the earth with high quality, open agent traces. no moats for the closed labs.

@ProgrammerDude Likely a miscategorization. Some app put something in ~/Library and now it’s misclassified.

140GB out of 256GB is used by "System Data" and "MacOS" and I have no idea how to shrink it. I can't even install the new MacOS Update because of it. What the fuck?

@KentonVarda Just keeps coming down. In 6mo time it may be halved, then halved again.

Wow, GLM-5.2 at home at 80tps for "only" $40,000. github.com/jamesob/local-… Obviously still out-of-reach for most, but getting close to being cost-effective for the tokenmaxxers. A year ago there would have been another digit on this number. Where will we be next year?

@halfpennymac hey guys, your product is sick. Where are you collocating in Oxfordshire? I have an idea I’d like to discuss with you. Couldn’t find a contact email so here I am 👊

@hthieblot Why raise if you’re doing $10M in revenue?

Series A expectations have gotten so insane that hundreds of great companies can’t raise despite growing 50–100% YoY on $5–10M in revenue. They’re not growing 300%, so traditional Series A firms pass. Seed funds are now leading the Series A themselves. Big opportunity

@ivanburazin Agree for a complex novel product for Daytona, disagree for an already-commoditized market such as Hetzner, OVH, EC2 etc.

I don't know why people think "we're cheaper" still works in the big 2026. I talked to three startups last week, and all three led with pricing. 30% cheaper, 2x less than [incumbent in their space], we undercut everyone, etc. The moment price is your selling point, you've commoditized yourself. Now, the only thing that matters is the number. And only commodities with great brands make real money. Evian sells the same water. You pay more because of the brand perception. But most startups can never compete on that. So if you're pitching cheaper, you're driving your company towards a dead end.

Taking the bait. You’ll have income tax at roughly 40% bringing your nominal yield down to 3.6%, then inflation running around 2-4%, meaning you’re basically treating water. If you invest less, don’t make much other money that year, you could squeeze out 1-2% of real returns with an effective tax rate of under 25%. But either way inflation is eating more than half of that 6% from the bottom, tax eating up to 40% of it from the top.

If you pop 10m into an account that’s $600k a year to live off of. Anything you save you get to throw back in and grow. This is wild if real.

@jaydrogers Dedicated?

Holy smokes! A 2 CPU, 2 GB server with Hetzner now costs ~$20/mo in the US. Just earlier this year it was about $5/mo This is why I love running a VPS. You can run Linux anywhere so you're never vendor locked. It's time to re-evaluate the market again😅

@kuberdenis Unless that subscription is subsidized currently right?

Assuming I need $7000 to get a rig of 3090s to efficiently run GLM-5.2 locally, this would translate to ~5 years of a frontier Max 5x subscription. Local models are a fun, futuristic dream, but nothing else. Sad.

You should consider buying them and putting them in colocation. Enormous savings, you’re paying for two margins with your current setup (bare metal provider + their colocation). @topflightpc can build them for you and ship them straight to colocation letting you capture the savings without having to get your hands dirty.

we started renting big bare metal servers and slicing them up into VMs for each person on our team this is basically the setup i've personally used for years now - esp useful with an opencode server running in there if this goes well we'll make this a public product