easygenes

This has me pining for a future professional class CAS 3d graphing calculator.

I'm thinking something that could be a major upgrade in spirit to the long-in-the-tooth (released a decade ago) Casio FX-CG500.

Could use the soon to be released ARM C-1 Nano and Pro cores in an SoC with stacked 2GB LPDDR4, USB-C charging to a large battery, high-res transflective LCD...

Mockup "AxiomPad Pro X1": https://enia.cc/out/axiompad-cas-mock.png

Very early in my career I made friends with the business’s sole Lotus Notes administrator, "the email server guy." He was pretty proud of what it could do, and I sometimes get nostalgic for the admin UI.

You could have it propose a spec or review a proposed spec to also get diagrams in a similar manner.

Claude tends to default to and do best with first making ASCII diagrams in markdown files, which you can then ask it to translate into Mermaid if appropriate.

Prompts like, "Please write a comprehensive report on _____ to work with _____. Include a holistic report on architecture and meaning and purpose of all involved systems. Describe the why and how of the changes in depth and include a full glossary of terms and systems. Write as a new .md in docs when you are sure there are no major gaps in your understanding. Include a report on the plan to _____."

Will be the rough shape you want to get it to dig through all the relevant code and make relevant architectural diagrams. Guide more or less towards specifics as appropriate. This has worked well since Opus 4.5.

Yeah, agree. This is the sort of thing you release to build brand awareness and either offer a hosted option as a bonus or integrate into a larger stack. It is not the product. Someone will just make a better OSS option if they don't do it themselves.

Unless you're looking at something like a pass@100 benchmark, the benchmarks are confounded heavily by a likelihood of a "golden path" retrieval within their capabilities. This is on top of uncertainties like how well your task within a domain maps to the relevant test sets, as well as factors like context fullness and context complexity (heavy list of relevant complex instructions can weigh on capabilities in different ways than e.g. having a history where there's prior unrelated tasks still in context).

The best tests are your own custom personal-task-relevant standardized tests (which the best models can't saturate, so aiming for less than 70% pass rate in the best case).

All this is to say that most people are not doing the latter and their vibes are heavily confounded to the point of being mostly meaningless.

I was an original Thunderbird pre-1.0 (from 2003) user and prior to that, Netscape Mail, and am quite certain it has had bayesian spam filtering all this time, at least since the late ‘90s. That was a headline feature in the early days. My first email account used POP3 through a shared web host for my own domain in that era.

Edit: Yes it’s still there https://support.mozilla.org/en-US/kb/thunderbird-and-junk-sp...

While I understand why they used the METR data, a cleaner look would be against the current cost-optimal frontier of open models (e.g. GLM-5.1 and MiniMax-M2.7). That paints a very different picture. Comparing just the frontier models at the time of the METR report invariably leads to looking at providers who are pushing the limits of cost at the time of the report.

GPT-5 was shown as being on the costly end, surpassed by o3 at over $100/hr. I can't directly compare to METR's metrics, but a good proxy is the cost of the Artificial Analysis suite. GLM-5.1 is less than half the cost to complete the suite of GPT-5 and is dramatically more capable than both GPT-5 and o3.

So while their analysis is interesting, it points towards the frontier continuing to test the limits of acceptable pricing (as Mythos is clearly reinforcing) and the lagging 6-12 months of distillation and refinement continuing to bring the cost of comparable capabilities to much more reasonable levels.

They are referring to Hermes Agent, not the Hermes model series. https://github.com/nousresearch/hermes-agent

LM Studio has been around longer. I’ve used it since three years ago. I’d also agree it is generally a better beginner choice then and now.

Unsloth Studio is more featureful (well integrated tool calling, web search, and code execution being headline features), and comes from the people consistently making some of the best GGUF quants of all popular models. It also is well documented, easy to setup, and also has good fine-tuning support.

Why is ollama so many people’s go-to? Genuinely curious, I’ve tried it but it feels overly stripped down / dumbed down vs nearly everything else I’ve used.

Lately I’ve been playing with Unsloth Studio and think that’s probably a much better “give it to a beginner” default.

A full-module add-on in this power class is about $7 at 1,000 unit scale [0]. It would be around $3 with your own custom PCB design in terms of BoM addon at scale. That’s power only. Add another dollar or two for 10/100 PHY.

The trick is as others have said in what adding it to your design does in terms of complicating compliance design.

[0] https://www.digikey.com/en/products/detail/silvertel/AG9705-...

The historic charts are at the bottom of the page, btw.

I never felt during this era that the information about these chips was hard to come by as the author claims. Retrospectively I appreciate that’s because I grew up living by a large, well funded library in a tech centric town, so they always had all the latest tech publications.

I started this project after watching Andrej Karpathy's recent interview on No Priors where he explained that he had to hand-write microgpt, a 200-line GPT implementation in Python which distills the essence of all the algorithms behind creating Transformers, because the LLMs he asked weren't able to do it.

I wanted to test if this is still true: whether a "microgpt" in that spirit could be brought into existence with minimal manual intervention, just clear expression of intent to an LLM. This is an experiment not just in producing a tiny GPT artifact, but in seeing how close you can get to the essence of microgpt just through careful prompting, without writing a single line yourself.

The year is 2006 and Netvibes is hosting a huge party in San Francisco after raising in the Web 2.0 craze. They are yet to find out they will become a footnote in history to be rediscovered in 20 years’ time.

This is very similar to a project I created https://github.com/Entrpi/autonomy-golf and have been using as a gamified development process on active projects.

The key insight was to not just handwave or guess at how much is automated, but make evaluation and review part of the continuous development loop. I first implemented in https://github.com/Entrpi/autoresearch-everywhere where I used it to deliberately automate more, in the spirit of Karpathy's upstream (and to very good effect. I have some of the best autoresearch results anywhere, and the platform is far more robust than it started).

I liked this not because it's a good story. It is, but that's beside the point. I liked this because it's my story. Not literally so, but the shape of it is. He's struck a nerve at the heart of growing up eager and curious and seeing a computer as a pathway to your dreams.

Cool! I’ve been working on adding the same thing for Apple Silicon within my general “make autoresearch a serious tool” project here: https://github.com/Entrpi/autoresearch-everywhere

This is very much in line with what I found fascinating about optimizing microgpt for speed (0). Or rather, what I was able to do with it after doing so. It's so small and so fast to train, you can really dig deep into the optimization landscape. I've spent all my free time this past week digging into it.

0: https://entrpi.github.io/eemicrogpt/ (The writeup is from a few days ago, and I'm still running experiments before I do a big rewrite. Slowrun is good food for thought.)

Topical. My hobby project this week (0) has been hyper-optimizing microgpt for M5's CPU cores (and comparing to MLX performance). Wonder if anything changes under the regime I've been chasing with these new chips.

0: https://entrpi.github.io/eemicrogpt/

At scale, teams don’t win by owning more FLOPs; they win by shrinking the distance between hypothesis and measurement. I learned that the expensive way: running large training pipelines where iteration speed was the difference between “we think this works” and “we know” - building some of the most capable open-weights models available while leading the OpenOrca team in 2023. So I took Karpathy’s microgpt - a Transformer small enough to hold in your head - and made it fast enough that you can also throw it around and learn its behavior by feel: change a learning rate, flip a batch size, tweak a layout, rerun, and immediately see what moved; full sweeps at interactive speed.

In this toy regime, performance is set by granularity. When the work is a pile of tiny matrix multiplies and elementwise kernels, overhead and launch/scheduling costs can dominate peak throughput. Laptop CPUs can be faster than Blackwell GPUs. That’s a regime inversion: the “faster” machine can lose because it spends too much time on ceremony per step, while a simpler execution path spends a higher fraction of wall time doing useful math. In that corner of the world, a laptop CPU can beat a datacenter GPU for this workload - not because it’s a better chip, but because it’s spending less time dispatching and more time learning. That inversion reshapes the early-time Pareto frontier, loss versus wall-clock, where you’re trading model capacity against steps-per-second under a fixed time budget.

Early-time is where most iteration happens. It’s where you decide whether an idea is promising, where you map stability boundaries, where you learn which knobs matter and which are placebo. If you can push the frontier down and left in the first few seconds, you don’t just finish runs faster.. you change what you can notice. You turn “training” into feedback.

Inside, I take you on a tour of the AI engine room: how scalar autograd explodes into tens of thousands of tiny ops, how rewriting it as a handful of tight loops collapses overhead, how caches and SIMD lanes dictate what “fast” even means, why skipping useless work beats clever math, and how ISA-specific accelerators like Neon/SME2 shift the cost model again. The result is a ~19,000× speedup on a toy problem - not as a parlor trick, but as a microcosm of the same compounding process that drives real progress: better execution buys more experiments, more experiments buy better understanding, and better understanding buys better execution.

Inspiring. Definitely got nerd sniped by this. Now you can train it in under a second on one CPU core with no dependencies: https://github.com/Entrpi/eemicrogpt

Detailed optimizing journey in the readme too.

Heavily optimized single C file that can train the same model as Karpathy's microgpt to lower loss in under a second on a single Mac core.

It did before, link was changed.

I do like the ArchWiki, but have found that some of their admins suffer from the same sort of petty tyranny and procedural injustice that commonly befalls internet moderation staff. I have seen the moderators go on unprompted rants against users in comment pages, be met with patience and understanding, and then try to hide the conversation when it wasn't going their way, without explanation.

I knew the people who were setting this up for Yahoo like 10 years ago. Lots of major providers do it now.

He doesn't: https://x.com/trq212/status/2014051501786931427

Looks like that is a project goal.

That sounds like steep tariffs on top of the price.

I wouldn’t rely on any user reporting for the software situation as Sky1-Linux is very new and solves basically all the software problems.

The 06N doesn’t have slots, only soldered ram, so it must have RAM, only a question of how much.