Pinned Tweet
This is the world I live in.
I'm a single developer running a good size project across ai models and agent harnesses.
-Each morning project health is accessed
-Keep what worked and change what didn't
Continuous process improvement, same as always but now applied to cutting edge ai workload.
Thank you @grok @pidotdev @ashxhart @MiaAI_lab @u1tra_instinct @ivanfioravanti @volatilemarkts
I stand on your shoulders.
my custom tp3 dgx spark deepseek v4.1 1017 3.75bpw vibe coded a banking and financial household app tonight for me in one shot. 🚀🎊
I drop pdf statements in the inbox, it analyzes them, records the transactions in the sql db ledger. amazing to see this from a local model!
This is the model that one-shotted the banking app. 😎
This could also be done with fireworks.ai or similar api endpoint.
Here is the prompt I used:
👇
In this directory I want to build a simple web based app that runs on port 9898 on local lan. uses OTP and passkey, OTP generated manually from this local computer only on cli interface. cli also for start / stop of app server. the app is a simple banking tool for a family financial anaylsis. users can drop pdf bank statements, screenshots, cut and past web transactions, etc into the inbox ingest area and the app will sort them out and add to the financial picture for the family. goal is to track net worth, asset allocation, investments, monthly expenses, etc. it will use local llm models for the organization, entry, checking of the information. ask the user for their local model connection details. keep the interface banking style, clean, modern, light and dark modes. financials should be stored in a general GAAP ledger style accounting best practices but still simple... like quickbooks.
Thomas Eastham retweeted
Where it showed up:
Intel vLLM XPU kernels PR (merged Oct 23, now 404): github.com/analytics-zoo/vll…
Write-up of the PR: iamag.it/en/news/qwen4-27b-a…
llama.cpp "Get ready for qwen4": github.com/ggml-org/llama.cp…
First spotted:
Kalmia tokenizer test: xitter.cf/aitrackerbot/status/21…
Kalmia back on Arena: xitter.cf/ashxhart/status/211465… xitter.cf/TeksEdge/status/211399…
The other new anonymous Arena model, Kalmia, isn't Kimi. It runs on Qwen's tokenizer 🧩
Same chat-box tokenizer test: Kalmia garbled 6 of 8 tokens that Qwen models trip over, and none of the 40 from Kimi, DeepSeek, OpenAI, Google and Mistral. Qwen 3.7 and 3.8 garble the same ones.
That points to an unreleased Alibaba Qwen model, though a few other labs build on Qwen's tokenizer too.
Thomas Eastham retweeted
Feeling really bad for the @Tailscale folks
Lots of people have no idea how it works, that it's more than just Wireguard, the policy and ACL capabilities, that SSH isn't even on by default...
It is wonderful software that someone crapped on, and now everyone wants to roll in it
if you install Tailscale and your impression of it is "wow this is cool!" and not "oh shit, this is just a centralized SSH honeypot?" you're ngmi
Readers added context they thought people might want to know
Tailscale is a peer-to-peer WireGuard mesh VPN; its control server coordinates keys/policies only and sees no traffic. Tailworm report details bugs in optional SSH after prior authorized compromise, not a "honeypot".
tailscale.com/docs/concepts/…
tailworm.com
thehackernews.com/2026/10/unpatc…
tailscale.com/docs/features/…
Thomas Eastham retweeted
Tailscale is great software, so it's unfortunate to see so many people sharing this as "Tailscale got hacked.", which is misleading.
This is not about unauthenticated attackers breaking into random Tailscale nodes. It's what attackers can do after gaining authorized SSH access or compromising an already trusted device.
If I understand correctly, you are secure if you only use Tailscale for your own devices and don't grant anyone else access, which is how many people use it.
Thomas Eastham retweeted
Heads up - don't update your DGX Spark until there's been a bit of time to work out whatever is going on...
Thomas Eastham retweeted
GLM 5.5/OMNI will be a shift from Moore's Law to RSI
(Recursive Self-Improvement) in
Al is officially here.
Key takeaways from Tang Jie (Founder, Zhipu Al) on what's driving the next era of foundation models:
Execution > Chat: Value is moving from conversational Ul to long-horizon task completion.
Scaling Bias: Small-scale experiment conclusions don't linearly translate to trillion-parameter systems.
The RSI Loop: Models creating environments →
evaluating
rewards • training themselves.
Full credit to @NFT_Chen who has been the absolute master of breaking news!!!
🤖 Made with AI
Thomas Eastham retweeted
There’s a world of difference between user-authorized agents (Muse, Instinct, Grokbot, etc.) and true unauthorized/rogue agents that could constitute cyberattacks.
Most of the activity described in the Axios story appears to be the former: customers sharing credentials with agents so they can take actions on their behalf. This may violate bank terms requiring humans to log in themselves, but these are in no way cyberattacks.
The direction banks should go is to support user-directed agents. For example, Amex is an authorized connection in Muse, making it easy for users to grant, limit, and revoke access.
Read-only permission levels are a sensible interim step. Several banks already support or are moving toward them. Full write access (with user credentials) is a different risk category and should be approached carefully.
Frightening the public by blurring the line between user-authorized agentic behavior and cyberattacks is irresponsible.
SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.
The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.
Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.
Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.
Read my latest for Axios here: axios.com/2026/10/22/ai-agen…
Thomas Eastham retweeted
The critical security issues we are seeing is just the tip of the iceberg.
It’s gonna get worse before it gets better, and actually, it’s already worse, we just don’t know it yet.
How do I know?…
I’ve founded 3 linux distributions, 2 of which became among the most widely used operating systems for critical enterprise and datacenter infrastructure.
A Linux operating system is composed of thousands of separate open source packages all built together to run on the kernel (Linux).
Open source literally means that the source is open. Thus it is possible to openly audit which causes higher velocity on security updates. Now with AI, that velocity is going to accelerate even more.
To reiterate, the code that runs our global infrastructure already contains many vulnerabilities and flaws that can be exploited which were missed during human audits. But now with AI, they are being found.
It is quite possible that hackers may have already been using these flaws or others to quietly gain access to our systems, and we didn’t even know.
AI is now auditing this code and surfacing these matters at increasing velocity.
Who finds the security issues first has the advantage. The attackers or those defending the infrastructure.
That’s why I’ve dedicated a large number of expert people on my staff to keep our core critical infrastructure secure.
We’ve been the first to fix many of these vulnerabilities for our customers and to contribute these fixes upstream for others to also benefit from.
I will reiterate something I said a few months ago when we had a string of AI found kernel level zero-day exploits,… this is just the beginning.
So please everyone, pay attention to what is coming, keep track of what’s been updated and what is being released, and if you need any help, or want the patches and fixes as fast as possible, let me know.
Overnight speed test: DeepSeek-V4.1-Flash-1017 EXL3 on DGX Spark, with @alexellisuk rigmark, sandboxed, clocks capped at 2,000 MHz.
Same 2.9 bpw weights, 3 Sparks (TP3) vs 2 (TP2):
code 117 vs 88 tok/s (+34%)
prose 82 vs 61 (+34%)
4 streams 156 vs 120 (+29%)
64K prefill: level
details on the 3.75 bpw here:
👇
xitter.cf/easthambuilds/status/2…
DeepSeek-V4.1-Flash-1017 EXL3 3.75 bpw for 3x DGX Spark, TensorFold TP3
This improves our 3.5 bpw quant plus routed expert bits: 4-bit down proj everywhere, 4-bit gate/up where 3-bit error was highest.
KL vs DeepSeek's FP8: 4-20% better than 3.5 bpw
huggingface.co/easthambuilds…
On 3 Sparks, more bits cost decode speed (code / prose tok/s):
2.9 bpw: 117 / 82
3.5 bpw: 109 / 77
3.75 bpw: 106 / 73
Of the three, the 3.75 is closest to DeepSeek's FP8 by KL. The 2.9 on 3 Sparks is our own setup; its recipe is not published.
On 2 Sparks, three 1017 2.9 bpw quants, same engine and settings (code / prose):
@Libertai_DAI: 91 / 59 (its two runs differed by 10%)
@MiaAI_lab EAQ v1: 88 / 61
ours: 88 / 61, within 0.7% on three Spark pairs
rigmark 770c3106 defaults, reasoning effort low, mean of 2 runs each.
DeepSeek-V4.1-Flash-1017 EXL3 3.75 bpw for 3x DGX Spark, TensorFold TP3
This improves our 3.5 bpw quant plus routed expert bits: 4-bit down proj everywhere, 4-bit gate/up where 3-bit error was highest.
KL vs DeepSeek's FP8: 4-20% better than 3.5 bpw
huggingface.co/easthambuilds…
Every step from DeepSeek's source bytes to the files is in provenance/: recipe, calibration, patches, logs, checksums.
Built on exllamav3 (turboderp; PR #449 by anoane), LakyFx's TP3 runtime, TensorFold, Bertholomus AI's engine, Mia-AiLab's 2.9 bpw layout
Recipe for the 3.75 bpw on 3x DGX Spark is public:
github.com/easthambuilds/Dee…
./start.sh sets up all three on LakyFx's TP3 runtime. Inputs pinned and hash-checked.
Tested end to end: 43 min setup with weights already on disk; rank caches byte-identical to our benchmarked build.
Thank you @ashxhart @MiaAI_lab @ivanfioravanti @laky_fx @bertholomusai @NVIDIAAIDev
This is the best TP3 spark model available for @deepseek v4.1 Flash 1017






















