zihotki

I wonder why there is no Github Copilot in the comparison. It also supports BYOK and is quite capable

I wonder what numbers you'd get using another system one model - Contrastive Language Model https://contrastive-lm.notion.site/

That model scales very well with quantities of requests.

Technically, anything what you plug into your car charger would do. The requirement is to have a licensed meter in the charging box. When setting up, you configure connection to your ERE provider and it just reports.

There are two parts in the data you supply to Jev for classification - the prompt describing your classification and the data. The data can be quite small - a simple chat message. And prompt part could be considerable since you need to describe your rubrics well.

With Jev you each time pay for your prompt, you can't cache it.

I would hold your horses to paint it as dirt cheap.. In my cases for spam detection Luna was 20% cheaper due to prompt caching, although not as fast.

That's what ADRs (architecture decision records) are for.

Documentation of the project are the projections of coded intent at a higher zoom level. C4model is an amazing concept and applies well to the context management and docs

In Netherlands you could also get paid for charging your car. There is such thing called ERE Certificates - emission reduction unit certificate is generated by offsetting emissions. Pollutters have to buy them to offset the harm. It's up to 0.1 euro per KWh (electricity costs 0.28 euro per KWh), depending on the demand.

And don't underestimate the complexity and importance of eval suite. It's a lot of work to get a good dataset. Your agents are as good as your dataset is.

My bet is that humans will kill humans faster than AI. AI enables not only good actors, but also bad actors: Russians, Iranians, Houthis, other terrorist groups - they all now effectively use AI to speed-up development of their weapons as well as hybrid warfare.

Landing page screams llm-isms, and after reading and looking around, I still can't identify what problems it tries to solve. The features described are just standard integration glue, they have little relevance to context or why's.

Why does it worth investing time into adopting this one instead building an in-house solution tailored to existing process?

Have you tried reading/reviewing the code claude generates? That could have prevented the issue with the accounts.

If everyone starts applying cooldowns, won't it postpone the problem? So now there is a considerable amount of users who are affected and someone from the affected group discovers the infection and reports it.

But if everyone will be delaying updates, won't be there less chances to catch it in time? I'm not fully sure if it's possible to preventively scan all NPM packages or how much compute it would require.

That's a good question indeed. Also I wonder why they picked java as the implementation language.

Since LLM can create the pages for non-techical people, it for sure can also help publishing the page

That sound interesting. Would you mind sharing it?

That resonates a lot with my experience in Netherlands. It's way friendlier for expats but the barrier is there

Similar - no (I'd be glad to be proven wrong), but you can find some small suppliers focused on specific niche like https://www.motedis.com if you need the profiles.

I'm building a homelab as well and fully second the ode. It feels incredible to do the setup via llm, like they are the perfect fit.

tldr - Qwen 3.5 9B finetuned on a "simple secret dataset" (500 million tokens of Claude 4.8 Opus, Claude 5 Fable, GPT 5.5 and similar models)

I found that a good tool helps a lot once I switched to Github Copilot app. It solved the friction and mental tax for me. I easily manage 4 sessions in parallel on same or different projects while 2 was max in the past. The bottleneck now is only in review and decision making.

Not that much - a bit better but still negligible. Prefill is highly parallelizable and benefits from many cores. NVidia GPU's simply have several times more cores than even Pros.

From personal experience - it works, but you won't get a comfortable time to first token (latency is high). The reason is that prefill on Macs is bad. You need to have a lot more cores to do it quick. It's close to instant for small models on NVidia GPU's but on Macs it takes a few seconds to get the answer for a simple prompt. And the time grows proportionally with your context size.

I wonder how it compares to https://github.com/refactoringhq/tolaria They look similarly scoped. I haven't used it yet and it's great that there is more choice now

In real life only a minority of teams achieved good results according to Circle CI report.

And there is one notable anecdotal source proving that it is possible - https://ideas.fin.ai/p/2x-nine-months-later

The solution is called curated package feed. Using private one instead of default public package source, you get a trusted source. All updates should be done by first pulling the new packages to the private feed and only after passing the reviews - update.

A little bit of bureaucracy in form of best security practices helps with supply chain attacks.

Are there any benchmarks/evals to back the claims? Or how do you know that it helps reducing waste?

Isn't that exactly the marketing part? He is the marketing and the channel, he has optimized marketing process.

And the best recommendation security teams can give - keep your SBOM strict, use min release age policy (sounds more like band-aid). That's a scary world to live in.

I wonder if adding an artificial barrier in form of a donation could help. That's probably the only remaining way to show the good faith.

Pencil and paper - that's the classical way a lot of developers from western europe were taught to code. Back to the basics