ggeorgovassilis

The guy that sells mainframes also gives away the Granite models for free. I've found them to be quite capable in content retrieval and long-running workflows, even (or especially?) the small ones.

I started using it last week when copilot started acting up against the Deepseek models. It's fast, solid, I love the architecture but be warned: it submits very very many input tokens against the LLM endpoint, so robust & cheap token caching is a must.

It also has a fairly decent vscode integration. I download & build it from the repo. I've submitted a few PRs to it.

https://github.com/Jabofish/deepseek-harness-vscode/

Author here. I'm running offline LLMs and using openrouter. I observed that many harnesses stumble over their communication with them: network errors, malformed tool calls, hallucinated response tags, empty responses etc. which leads to aborted flows and user-visible errors. I wrote llm-gauze as a transparent openAI compatible proxy that works around many of these issues.

If you have worked with open-weight/local LLMs, you've known both the highs and the lows. Lows there are many: malformed response tags, empty responses, stuck loops, overthinking, inference engine errors and much more. You don't notice those when using commercial models, because the vendors iron out those quirks in their gateways, but you're left standing when using free models. Gauze is a transparent openAI compatible gateway which monitors and corrects bad LLM responses.

> Seriously, if a single politician stepped forward and said "i'll bring down ram prices"

Or abolished VAT (the meaning of VAT is that you pay a "rent" for all the infrastructure used to produce the thing) and import taxes (protect your market) on stuff we don't produce in our markets anyway.

Anecdote: I operate a long-running (overnight) agentic workflow which scrapes marketplaces for prices. They run in VScode with several agents coordinating tasks between them. Agents run on a mix of open weight models (each to their strengths). The workflow used to crash after a few hours for different reasons, which always were rooted in an LLM error: either it would produce a wrong tool call, its context window would fill up, it would return only a thought stream and no response etc. Using Gauze, the workflow always runs to completion.

Wrote this for fun over a decade ago. Still around.

https://animasci.com/

Thanks. It's available only for the US layout, not UK.

I'm writing down a few impressions: - the layout is unusual, but I get the motivation. Distances are minimised and letters are arranged so that ambiguity is removed. - although I'm very slow, I haven't made a single mistake so far. Clearflow allows me to swipe much more accurately than stock gboard. - the square keyboard layout unfortunately means that half the letters are constantly hidden behind my thumb. As I'm unfamiliar with the layout, this means that before swiping a word, I have to look at the layout, memorise letter locations and plan the movement - since I write in multiple languages and Clearflow is available in only one of them, I would have to memorise a completely new layout for a language I write in only half the time.

I learned something. Thank you

If you don't mind, how do I get ClearFlow in gboard? GBoard on Android only has language selections, not keyboard selections.

On two levels: architecture and understanding. Architecture: I divide the solution components of my architecture into two groups: the ones where a security breach spills over their scope and the ones where it doesn't. For the first category (eg. network- or user-facing), dependencies will be limited as much as possible, meaning I'll forgo convenience and features. I'll pick LTS or older versions with no known vulnerabilities. The second category is locked up in containers with minimal connectivity, with on-demand run-time schedules. Understanding: depending on risk and importance, I actually check out a dependency's source code and have an AI review it. Then rebuild and self-host.

Edit: this approach sounds like it could be bundled into a couple of agents.

The Paul G. Allen School YouTube channel has over 10 years of scientific lectures on advanced research topics, but has only 24k subscribers.

I just watched Michael Abrash [1] (of Graphics Programming Black Book fame) talk about the history and future of VR.

https://www.youtube.com/watch?v=V5AwlNbAvCo

Many are playable with Steam on Linux, each game in the store states whether that's supported or not. Even non-supported games allow an override. I've tried that for a few with varying success. Steam has so far refunded purchases that didn't run on Linux. Then there's Lutris which runs many old games fine.

tbh you can go pretty far with just cloud VMs and virtual networks.

The company I currently work for employs several mid-level software developers who've never worked with "the cloud". Their focus is software design, not operations (which is what the cloud is mostly about). The contact point with operations is container technologies. Their applications do REST, connect to databases etc. like they would do in a bare metal data centre. I occasionally need to remind them of low instance uptime guarantees, the need to plan for horizontal scaling, data replication and disaster recovery (I basically have to design the application architecture for them) but apart from that there's no friction in the day-to-day work.

A piece of unsolicited career advice: at the work place, continue being enthusiastic about software technologies and keep opinions on infrastructure close to your vest.

I miss Google reader. Easily self-host the next best thing, FreshRSS, with docker, a web interface and secure HTTPS.

Will epochal advances in Artificial Intelligence complement human expertise, thereby increasing its value, or render it increasingly unnecessary? During both the industrial and computer revolutions, the forms of expertise rewarded by the labor market changed substantially, with vastly uneven consequences for workers in different occupations and possessing different education levels. These forces will play out differently again in the AI era. While the future is a design problem not a forecasting exercise, David Autor will discuss the challenges that AI poses for the labor market, as well as some of the novel opportunities it offers.

CES Munich Lectures in Economics 2025 Expertise, Artificial Intelligence, and the Work of the Future David Autor, Daniel (1972) and Gail Rubinfeld Professor, MIT Economics, USA

I came here just to complain about that :-) All LLMs I used seem to give more weight to things at the beginning of the context window and omit many details. Eg. I tried this simple thing: pasted a friend's and my CV into Gemini and asked it to recommend topics for a joint conference presentation. Results depended greatly on the order of CVs pasted in.

> they areare mystical or part of their mental work.

In ancient Greece, epilepsy was called the "holy disease" and it was believed that gods speak through the patient during a seizure.

2017

And a more digestible report: https://www.theregister.com/2025/10/29/ey_exposes_4tb_sql_da...

Interestingly, the X account is private now. That's an effective way to contain fallout.

Runs 24/7. Also, I already host other things on the server, so there is no added cost.

I'm self-hosting FreshRSS [1] with Docker on a Hetzner VM. Fast, clutter-free, has everything I need.

[1] https://freshrss.org/index.html

Edit: typo

I like thinking and blogging [1] about electronic circuits despite lack of formal education. It being a hobby, I must concede that my involvement and proficiency remain well below industry standard.

[1] https://blog.georgovassilis.com/category/electronics/

Unfortunately the installation experience on Android has been getting worse. In the beginning a website could prompt the user for installing the app offline, then it became a semi-heuristic thing when the browser would trigger that even, now you have to do it through an entry in settings. Also, if you've installed multiple browsers, it's a dice roll which browser will be used to launch the app.

ICA can isolate individual signals from a mix. The demo mixes three radio recordings together and then extracts them again.

I'm using RethinkDNS as a combined firewall and wireguard VPN. I added these two block rules:

127.0.0.0/8

::1/128

I'll update here with any issues.

Location: Athens, Greece

Remote: Yes

Willing to relocate: No

Résumé/CV: https://www.linkedin.com/in/georgovassilis/

Email: george at $lastname dot com

Industries: banking, biotech, consulting, finance, government, insurance, manufacturing, real estate, retail

Skills: agile, Archimate, business analysis, engineering management, enterprise architecture, ITIL, technical design, TOGAF, QA & test automation, UML

Technologies: Java, Python, Linux, Azure, HTML & JavaScript, SQL, Selenium, containers, various DevOps tools (eg. Jenkins, Terraform)

Wants: enterprise architecture/engineering/platform/testing IC or management