Pinned Tweet
Frontier-level AI on my desk, no limits.
JSpark3 runs GLM-5.3 Flash (abliterated) across 3 DGX Sparks. The model beats Opus 4.8 on DeepSWE and OSWorld.
Everyone said they'd be good once this existed. It's here. It only gets better.
GLM-5.3 Flash on JSpark3 v1.8 hits 73.3 tok/s on code and 49.1 tok/s on prose. Stock weights, pinned recipe, every number measured.
Recipe and card on Hugging Face:
huggingface.co/jakejharris/j…
When you think a few months ahead, and you consider that this is what cyber will need to defend against, we will all look back and wonder if Dario was right
I’ve noticed my codex usage has been getting drained significantly the past 24hrs relative to previous
I imagine @thsottiaux and co know about it considering they haven’t tweeted out their “improvement” of the day.
Running Tailscale SSH across your cluster? Read Tailworm.
A node already allowed to SSH in can ride unix socket forwarding to root on the target. Not an entry point, a lateral movement one.
Patch the second it ships, turn on Tailnet Lock.
Once you've done that... consider if Tailscale is truly needed or if it's just another attack surface that you need to consider.
thehackernews.com/2026/10/un…
Has anyone else’s Claude agents really been pushing you to sleep/get some rest? Can’t believe this rigamarole is starting again.
Black October is going to impact this election cycle. Whether anything serious actually happens, doesn't stop the US admin from saying that something serious is happening.
If Kimi 3.1 ablit hasn't been the cause of this CVE uptick, it soon will be. Those benchmarks are insane. This is just getting started.
Jake Harris retweeted
everyone should have access to cybersecurity models. the attackers already have them. there is no point in gate keeping. DEFENDERS ARE LOSING. RELEASE THE MODELS NOW.
Jake Harris retweeted
the beatings will continue until the labs give independent researchers easy access to SOTA uncensored models
Jake Harris retweeted
The market is severely underestimating compute demand over the next 20 to 30 years by at least 1,000x because financial models are treating AI data centers like standard enterprise cloud spending rather than a global utility expansion.
Moving from simple chat prompts to continuous, autonomous background reasoning and real time spatial computing will fundamentally break traditional grid connections, pushing total power requirements directly into a terawatt scale footprint.
Intelligence will be the largest market in the solar system.
The long term bottleneck won't just be semiconductor manufacturing or HBM yields, it will be securing dedicated behind the meter baseload energy like small modular reactors and direct to chip liquid cooling infrastructure to handle the thermal load.
Compute is the defining physical infrastructure asset of this century.
I'm also just looking out for the retail investors. 5 years ago Anthropic had just been founded. Claude 1 came out in 2023. Anyone who thinks they know what's possible 5 years from now, let alone 10, hasn't been paying attention to the last 5.
I’m just looking out for the retail investors. @anthropicai is not on track to have 1 drug in humans in 5 years and def not to cure most diseases in 5-10 years according to @DarioAmodei ‘s predictions.
This is only 3rd post about Anthropic wet lab “success” in 3 months (mini-binders, non-CRISPR and now non-anti-venom) in 2 months and they didn’t even use their prized wet lab (which is no different than my rented space); they just used CROs like anyone else.
We’re just now entering the next wave of open model releases, what are we most excited for?
0%GLM 5.5
0%Qwen 4
0%Something else
0 votes • Final results Jake Harris retweeted
Replying to @MrSage
It's a smart model.. I ran with huggingface.co/jakejharris/j… for about 24 hours of deep work and it killed it. Insanely good output.
Jake Harris retweeted
NVIDIA just shipped a 4B model that thinks in loops.
Nemotron Labs Dory 1.0 is on Hugging Face.
Same weights, you pick the compute:
1, 2 or 4 loops per token. No retrain. No model swap
Dive in
huggingface.co/nvidia/NVIDIA…
Has anyone out there made a benchmark for which models write the best prompts for other models? Promptbench anyone?
GR8 is the best way to run GLM-5.3 on TP4 right now:: github.com/kindlingai/gr8
Huge thanks to @mmastrac and team for building out GR8 and working with me to get JSPARK Ablit working with it. It's still an experimental build but if you've ever wanted an ablit steering config instead of seperate weights, you should check out:: huggingface.co/jakejharris/j…
JSPARK Ablit is a 1.18 MB add-on that makes GLM-5.3 much less likely to refuse a question.
Typical ablits are separate, permanently altered models. Ours loads into GR8 and switches on per request with no speed loss. It answered 90.6% of prompts the stock model refused in our pilot.
You can try the experimental release below 😎
huggingface.co/jakejharris/j…
EXLR8 wins on more than one axis - it's designed to be loaded by _any_ engine (reference loader provided!).
It's designed to be cheaply sharded (TP=1/2/3/4...N).
It's designed to load _fast_. A compliant loader can boot in under 30s.
It's sliceable and moddable by anyone. Don't like our weights? Make your own! You only need to fork the weights and publish a manifest (MB, not GB).
You can even add a steering vector (@jakeharrisdev's awesome abliterated GLM 5.3 is our first!).
It's not just a fast and small weight format. We went back to basic principles and re-engineered the whole weight format to be modern, composable, sliceable, pluggable and (best of all) as fast as physically possible.
It's not EXLR8 vs everyone else. It's the modern world versus a bag of safetensors.













