AI enabled mindset with passion for health and wellness projects.
Earth.Planet
Joined December 2023
- Tweets638
- Following657
- Followers168
- Likes1K
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.
Thomas Eastham retweeted
I’m no longer prompting Claude through chat, except for initial messages. The reason: chat has become unmanageable. Instead I add an annotate-anything layer to the apps I build. Each message is a linked markdown file (git committed). A coordinator watches them and assigns tasks.
I think we're going to see a new wave of interest in usability.
Chat is hard on the brain. Total opposite of what Don Norman wrote about. 100% holding in your head, remembering things, special modes, etc.
Nobody wants the stuff in their kitchen to be in a different place each time they walk in.
System of record + recognizable display are not going away. What's new is how fluidly we can change state WITHIN those as users.
Thomas Eastham retweeted
GPUs should qualify under the second amendment
Thomas Eastham retweeted
what if employees at labs haven’t actually unlocked some order of magnitude better way of working with 100 agents running around the clock?
what if their setup is just a function of not being token constrained?
their product velocity and quality do not suggest they’re living in some different reality. their incentive is to convince you they are. their valuations depend on it.
Thomas Eastham retweeted
Solve your own problems first.
If you don't even use it yourself, then don't expect anyone else to.
Tools I have built to save me (and others) money/time
- Agent Zero (Xero replacement), with MCP for my agents to do my bookkeeping/accounting.
- LocalRouter.app, full AI stack running locally on my own hardware.
- Autovisuals.io helps me decide what wheels/mods I want to make on my cars.
- Go Fast Stack (Drop-in AI-first software template)
- Lockpick.app helps me keep my SaaS products safe and secure
- Davit (Boat.dev but on your own hardware)
- simplifiedpdf.com/ sign/edit PDFs for free and totally client-side
- simplifiedimg.com/ edit/watermark images for free and totally client-side
- simplifiedqr.com/ QR code tool for free and totally client-side
- Buildthings.co AI/Software/Hardware company and most of my day jobs.
I have built many other products over my career that have had different levels of success. None of them has gone mega viral, but they all solved problems for me.
Even if they don't make me money, they link back to other ways I can potentially make money and solidify my reputation in the industry.
Thomas Eastham retweeted
The last time a new machine threatened a third of American jobs, we reinvented how we educated our kids.
That bet sparked an epic 100-year run.
It is time to make it again.
Joe Liemandt of Alpha School on why AI is today's 1910 tractor moment, and why parents once again have to rethink their kids' future:
"In 1910, 90% of America had an 8th grade education. There were basically no high schools. And over 30% of America is a farmer. So everybody's on the farm, 8th grade was enough."
"Then in 1910, the Case tractor rolled up on the American farm. You can go look at the ads, it literally looks like an Anthropic ad. 'Does the work of 50 programmers. Does the work of 50 farm hands.'"
"Literally everybody in America realized 30% of the jobs are going away. That path is gone, and we have to get our kids ready for this new future. And it was that same recognition I think that parents have today."
"What America did back then... they voted to fund an education that they never experienced. They knew they had to do something different than their background, see the world differently, to get their kids ready and get them the good jobs."
"That school system became the industrial system we've had the last 100 years. It served us really well for 100 years, but now it's time for the next change."
@jliemandt
Thomas Eastham retweeted
This is fascinating. As it turns out, when you stop making benchmarks public, open models perform significantly worse.
For me, Deepseek 4.1 Flash and GLM 5.3 have been an incredible revelation, these models are sooo good.
But when these labs/people start trying to sell me the idea that we have "Fable at home" and that "China caught up" I just can't help but roll my eyes.
I want there to be Opus 5.5 at home so badly, but the narrative spinning is so unhelpful. We've gotten to a point where we basically ignore almost every benchmark.
Really hope this is a new leaf where we can start to put a little more stock in the results.
Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets
The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks reflecting real-world usage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.
Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.
New to the Intelligence Index:
➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change
➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment
Upgraded:
➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions
➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised
➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool
Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0
Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1
Key results:
➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50
➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)
➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04
Thomas Eastham retweeted
Nvidia gave some DGX Spark love today. It profiled Link 3.0 Flash (128B model) and improved decode.
37.8 → 53.9 tps.
No hardware upgrade. Thanks, Nvidia!
It stacked three software optimizations
⚙️ Humming MXFP4 MoE kernels
🧮 FP8 LM-head quantization
🚀 built-in MTP/NEXTN speculative decoding
On the full 1,319-question GSM8K evaluation, it improved
Decode
37.81 → 53.89 tps 42.5% speedup.
Mean latency
15.21s → 11.49s
Accuracy
96.89% → 96.82%
So basically the same measured accuracy while getting a lot more decode from the same hardware.
And the synthetic test gets even crazier:
36.09 → 102.88 tps
Link in ALT
ALT https://developer.nvidia.com/blog/3-ways-to-boost-moe-inference-on-nvidia-dgx-spark/
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: nitter.gq/aitrackerbot/status/21…
Kalmia back on Arena: nitter.gq/ashxhart/status/211465… nitter.gq/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.