Nextify Now retweeted
today i learned there's a mystery tech founder who's spent $100M+ building permanent underwater human settlements
this summer, his company DEEP lowered its first underwater home onto a reef in Florida, 56 feet down
and the first crew of 4 is about to move in for a week at a time
here's what life down there looks like:
> they eat, sleep and work in a living chamber on the seafloor
> there's a pool of open ocean in the floor called a moon pool, so they can step out of the house and straight onto the reef
> a buoy on the surface pipes down air and power through a cable
> because the air inside is pumped to the same pressure as the water, they can dive all day without swimming back up to the surface
and that's just the starter version
next they're building one that holds 6 people 650 feet down for a month, with bedrooms that have giant round windows looking out into the ocean
their motto is literally "make humans aquatic"
and nobody outside the company even knows the founder's name
interestingly people actually tried this before, but it ended badly:
> in the 1960s, Jacques Cousteau and the US Navy raced to build homes on the seafloor, the same way everyone was racing to the moon
> the Navy even had an astronaut live underwater off California for 30 days
> then in 1969, a Navy diver died 610 feet down trying to fix the deepest one
> the program got cancelled, and almost nobody tried again
until this mystery founder showed up
can anyone guess who he is?
Nextify Now retweeted
RAM prices are so high that Red Hat engineers are looking at SSDs to hold part of your virtual machines' memory.
At KVM Forum 2026 in Vancouver this week, Red Hat virtualization engineer Peter Xu gave a talk called "QEMU memory tiering". The idea: keep a VM's busy memory pages in DRAM and move the idle ones out to cheaper NVMe SSDs, without the guest slowing down.
"As memory is getting much more expensive nowadays due to AI deployments, we have been looking into how to save RAM without sacrificing VM performance," the talk abstract says.
The simple route uses what Linux already does well: swap, plus per-VM memory cgroups. The catch is that QEMU doesn't know which pages live in which tier, so a live migration can pull swapped-out memory back in and badly hurt performance.
Three options were on the table: a new system call for tier-aware memory access, having QEMU run all the tiering itself, or the current favourite, existing kernel interfaces (mincore() + MADV_PAGEOUT) with small QEMU changes. A proposed kernel addition, MADV_PAGEOUT_FULL, would fix the last big problem: swap overshoot caused by async reclaim.
Nextify Now retweeted
Replying to @Tristan5d @grok
42 is the answer to life the universe and everything.
Nextify Now retweeted
inside, outside, through, beyond
Nextify Now retweeted
If you were over the weather; you’d be in space. 😑
Nextify Now retweeted
Our investigations into the recent wave of AI inference infra attacks point to a concerning pattern.
We believe Anthropic’s crackdown on distillation may be accelerating a shift toward stolen credentials and compromised inference infrastructure.
Our findings suggest certain Chinese operators are involved in networks reselling frontier model access through heavily discounted AI gateways.
The economics are simple: steal API keys, hijack inference capacity, and resell tokens at a fraction of legitimate prices.
We suspect these are not isolated incidents, but part of a growing underground economy built around stolen AI compute. Inference infrastructure is becoming a prime target for organized cybercrime.
Nextify Now retweeted
openai's head of design's website omg iansilber.com/ 😭
Nextify Now retweeted
We're releasing new Qwen3.8-27B GGUFs with Unsloth Dynamic V3.1! Expect 5-10% accuracy gains on the same size vs UD3 and other quants.
We do better vs GSQ-RCO, Exllama-v3 on the pareto frontier for Terminal Bench, DeepSWE and more. Quants above 2-bit to 6-bit are updated.
We did not update 3-bit quants from UD3 to UD3.1 as we found UD3 to perform on par with 3.1.
We also decided to benchmark Exllama3 quants as the community requested them. Exllama3 keeps the token embedding layer in BF16, which is different from ISTA and UD3 / UD3.1 and GGUFs.
We decided to exclude the token embedding from all plots. This does not mean the RAM and disk space usage of EXL3 is less - it's in fact more if you include the BF16 embedding layer. We provide plots for actual RAM+VRAM usage at unsloth.ai/docs/models/qwen3…, and it shows exl3 quants doing much worse.
For LiveCodeBench, GSQ-RCO showed Unsloth UD3 quants doing consistently worse than ISTA's. However we found the opposite - UD3 quants (our OLD quants) do better than ISTA on LCB on all sizes except the smallest. UD3.1 further improves on LCB as well.
We approximate that GSQ-RCO used a context budget of 96K, as that best fits the data for the runs we have. At 96K, GSQ-RCO does somewhat do better than UD3, but at 128K, UD3 does much better (and UD3.1 even more)
For ISTA GSQ-RCO's AIME25 benchmarks, their quants reach 100%, 100%, 100% and 96.7% accuracy recovery.
We estimate ISTA only did 1 trial - we did 14 trials, and we show on average the recovery rae is 93% to 96%. And in fact our UD3 old quants do better on average than their quants on 14 runs vs 1 run of theirs.
We are still improving our algorithms to make smaller quants and larger quants (and also 3-bit) much more accurate, and will definitely apply these methods to Qwen4!
All blog details at unsloth.ai/docs/models/qwen3… and the new quants are at huggingface.co/unsloth/Qwen3…
Today, we're releasing the highest accuracy Qwen3.8-27B GGUFs.
Our new Dynamic V3.1 GGUFs lead by 5–10% on average across Terminal Bench, DeepSWE & more. We’re working on even better version for Qwen4
Run on 12GB RAM.
GGUF: huggingface.co/unsloth/Qwen3…
Blog: unsloth.ai/docs/models/qwen3…

































