mtkd

Isn't the point of Jev that it generalises better?

It's a fast classifier you can use out-the-box, ~1.5bn tokens is about $40 (I've been hammering it)

It just works ... a whole bunch of low-level/low-importance workflow stuff that was getting farmed out to small/fast LLM models now has a competitive alternative ... and bits that hadn't even been considered to go into some external descision/classifier service can be tested/deployed at ~$0.00003/req

I don't get this wall of negativity on it, it's genuinely innovative/useful tech ... would expect HN to be more positive, regardless of whether it's the absolute best execution

I was a bit skeptical when read the initial pr on it, yesterday ran a test involving ~250M tokens, something we measure went from ~60% to >80% success (with almost no tuning) and at less than 50% cost the low-end LLM was running at, looking at it more seriously now ... the servers are US-only currently I understand and ZDR is by request

>It is not a big deal

have you ever held and read an old book?

The world you grew up in no longer exists

No, you get a AUP violation and have to manually swap the model

(I had same issue, just asked it to check some code that 4.8 had modified earlier in day)

Getting an agent working via an existing Slack setup is fairly effortless and the control over output format is useful

I had a look at getting same agent up in WhatsApp but it seems to need FB business acc to even start process, to get FB business you need an FB personal etc. ... looked like too much effort

Heating oil in UK from one of main suppliers, for delivery next week, has gone from 60p/L to 130p+/L in last 7 days

You can still look at almost any codebase and ask 'why is this bit not using a state machine here?' ... the AASM repo readme is very accessible even if you don't know Ruby: https://github.com/aasm/aasm

The link is a sales pitch for some tech that uses MCPs ... see the platform overview on the product top menu

Because MCPs solve the exact issue the whole post is about

But a small collective running that box, especially spanning timezones, could potentially be a viable alternative or will be soon -- with obv privacy gains too

They need to max valuation before hardware catches up and qwen3-coder can be run locally for free

It's easy to forget the product Anthropic are selling here, and throttling, is based on data they mostly pay little or no content fee for

Conversely, it's useful to get an immediate answer sometimes

6 months ago, "what temp is pork safe at?" was a few clicks, long SEO optimised blog post answers and usually all in F not C ... despite Google knowing location ... I used it as an example at the time of 'how hard can this be?'

First sentance of Google AI response right now: "Pork is safe to eat when cooked to an internal temperature of 145°F (63°C)"

>it's your problem to deal with to a degree

How is it not the responsibility of senior management at a major retailer to ensure an exploit at a vendor can't take the whole house of cards down?

Many other major enterprise clients out there are all over vendor security/compliance ... auditing and reauditing vendors to minimise chance of this happening or worst-case, if does happen, containing it and recoverying quickly

For the trivial cases that's fine (just using LLM does same)

But this particular project is not like a standard site and the CSS is in small fragments across 100s files and uses constants for some things like color values in places too

In that Loopple example you can see the conversion uses the Tailwind arbitrary value notation, the -[], so background-color:#afa8af gets converted to bg-[#afa8af], but I wanted nearest pure tailwind class bg-zinc-400, the agent seems to work out color distance fine so does all that in one-shot too

Last week stripped out all CSS from a fairly substantial project and replaced with Tailwind equivs, it got all but a few cases right

That was gemini-cli, I could see some mistakes on trial run so created a GEMINI.md with system prompt and project description (about 50 lines) which clarified some tricky source layout situations

Second run it was fine, ran for about an hour or so -- I had attempted to do it manually a while back but it started to look like it would take a week or two

I expect in a few days there will be a new tool launched that returns word frequency/velocity in recent biomedical papers ... so next year's PhDs can level things using an MCP function

Same was said about dejanews, stackoverflow etc. and intellisense

Unnecessarily critical take on a quality write-up

Much of the criticism of AI on HN feels driven by devs who have not fully ingested what is going with MCP, tools etc. right now as not looked deeper than making API calls to an LLM

In 2023 Huawei surprised with the Kirin 9000S in the Mate 60, this seems to get forgotten when talking about GPU moats and sanction effectiveness

It's moving very fast:

https://elevenlabs.io/

https://layercode.com/ (https://x.com/uselayercode has demos)

Have you used the live mode on the Gemini App (or stream on AI Studio)?

That's likely because OG devs have been seeing the hallucination stuff, unpredicability etc. and questioning how that fits with their carefully curated perfect system

What blocked me initially was watching NDA'd demos a year or two back from a couple of big software vendors on how Agents were going to transform enterprise ... what they were showing was a complete non-starter to anyone who had worked in a corporate because of security, compliance, HR, silos etc. so I dismissed it

This MCP stuff solves that, it gives you (the enterprise) control in your own walled garden, whilst getting the gains from LLMs, voice etc. ... the sum of the parts is massive

It more likely wraps existing apps than integrates directly with them, the legacy systems becoming data or function providers (I know you've heard that before ... but so far this feels different when you work with it)

It's very different to blockchain hype

I had similar skepticism initially, but I would recommend you dip toe in water on it before making judgement

The conversational/voice AI tech now dropping + the current LLMs + MCP/tools/functions to mix in vendor APIs and private data/services etc. really feels like a new frontier

It's not 100% but it's close enough for a lot of usecases now and going to change a lot of ways we build apps going forward

SaaS companies tended to need material engineering resource due to the software stacks and squad-style team structures in place -- also leaned heavily on costly metered infra

I'm not seeing anything like the same level of heads or stack complexity in this wave (Vercel, Firebase etc.), and the vendors involved get cheaper every day ... along with increasing ability to run models locally with no metered costs at all

It's deeper than the security issue

You could have two different packages in a build doing similar things -- one uses less memory but is slower to compute than the other -- so used selectively by scenario from previous experience in production

If someone unfamiliar with the build makes a change and the assistant swaps the package used in the change -- which goes unnoticed as the package itself is already visible and the naming is only slightly different, it's easy to see how surprises can happen

(I've seen o3 do this every time the prompt was re-run in this situation)

Rediscovered a wheel of gouda from covid prepping era the other day in pantry

Took my chances and broke the wax, despite advice from internet

Brittle but hard to cut -- 10/10, highly recommend putting one under a shelf for a few years

Migration to fly.io is simple enough ... it is much closer to the original Heroku both technically and as company (if you ever need to contact them)

That is different things though -- you can replace functions with AI tech without needing AI to write all the code and that is happening in a big way right now -- Krishna was talking about reducing customer facing roles by around 30% over 5 years in that earlier statement

As of right now the code assistants mostly just make existing coders more productive -- predicting 20 or 30% of new code in near term will be generated is not unreasonable -- 90% is a stretch as has been discussed here many times

Do really think LLM vendors that download 80TB+ of data over torrents are going to be labeling their crawler agents correctly and running them out of known datacenters?

One of risks of AI code assistance is that they are not necessarily looking at the wider picture when it comes to libraries used on a large code base

I was testing o3 recently and it kept changing the library used by a block of code every time it tried to fix an issue in the block that was unrelated to the library used (haven't seen that happen with Sonnet)

Easy to see how issues could creep in because a modification is made that switches to an inferior library/gem that exists in the code base or standard library so still passes tests etc. but doesn't need a Gemfile change