Roark66

Exactly, these general scores are pointless. Let's say the model is amazing on A but crap on B and C. It gets low score, but you still would benefit greatly if you need it for A.

I don't quite believe the benchmarks, especially supplied by the company.

Qwen3.8-Flash-Next consistently benchmarks better than both GLM5.3-Flash and the previous DeepSeek v4.1 Flash 0731. The DS 0731 was so atrocious it was getting beaten by small Nemotron in my coding tests (subset of tasks from open source coding benchmarks).

Now, new DeepSeek V4.1 Flash 1017 really gets the numbers they promised for the first one, but it's too big for me to run the 4 bit quant locally at interactive speeds.

So I kept Qwen3.8-Flash-Next as my go to.

I wish they released a flash variant that is no more than 200B (120B preferred). Something to compete with Qwen3.8-Flash-Next - my favourite local model.

Exactly, using 40k of your context upon startup is bonkers.

You know what. When Gemini 3.1 Pro Preview came out it was better than Opus at the time. It was my main model for 3 weeks.

Until it got lobotomised. Then it lost most of it's capability.

Similar situation happened to Anthropic's Fable. It was an amazing model until Opus 5.5 came out.

So now, I'm not even looking at a "new" model until it is a month or two old. By then we will know if they will destroy ability to make it cheaper to host.

The 2nd amendment is an absolute and utter bullshit against tyranny. In fact it is a method to make ordinary people not engage in peaceful protest when it is needed most.

Why do you need to protest at the slightest sign of the government becoming authoritarian when you have "the ultimate argument of force" on your side?

But it's an illusion. Guns are a nice hobby I enjoy as well. They are great for self defence, but good luck to you if you think you'll use it "against government tyranny"... You'll be dead and no one will care.

2nd amendment makes an ultimate duty of a citizen in a democratic country - to not allow it to turn into a tyranny, into "you don't need to worry about politics, worse case scenario you have your AR15 and a box of ammo".

While in a country without such false ideas people get out and protest in mass numbers.

I don't think I've ever heard of a proper protest happening in USA. The kind of protest that involves 10% of adults at least and happens in every city and town. (Also I'm not talking about rioting like in France)

Burning down a school to protest against class overcrowding is like something straight from idiocracy.

Peaceful protests work on governments, even those hell bent on authoritarianism, but they have to be massive.

Let's say you have a million or two of people show up in Washington. Will it make any difference? Nope. Why? Because it's under half a percent of US population.

Let's now say 10% of people in every city and town go to their town centre and protest. Does this make a difference? Yes it does, because it shows real power to everyone.

This should be obvious to anyone that has seen the videos from this incident. In any normal civilised country 10% of adult population would protest something like this next day in every city and town. Then those "agents" would be arrested and they would have a very public court case.

But in US human life is not worth much.

Yes, these are my colleagues. Same crap as before. At least now I can setup an automatic ai review that is told what is not allowed (the errors they do all the time and refuse to learn) and I get PRs that have passed 8 review turns and are somewhat better.

Actually more AI agents do the work faster and cheaper.

I use such orchestration for those 2 reasons. One, a supervisor model like Opus 5.5 (or fable before they lobotomised it) limited to 200k of context and told to keep it's progress updated in a text file is smart enough to understand the whole task.

It farms smaller sub tasks to a fleet of 7 Qwen3.8-Flash-Next agents. Not only it uses up 10% of the tokens opus would use those 7 agents run at an aggregate rate of 200tok/s. While opus would run at few dozen at most.

The Qwen in this example runs locally. If you don't mind sending your code to China you can spin a bunch of DeepSeek Flash agents and each will run at ~150-200t/s. Then your task gets done way, way faster

When running benchmarks mind how some providers lobotomise some models. For example DeepSeek Flash is fine when run via DeepSeek. The same model via Deepinfra and open router (despite setting routing to "exact") was far worse.

Regarding Qwen3.8-Flash-Next only Alibaba matches the quality I get when run local inference.

Mind I run mxfp4 and int8 kv (before I run unsloth's q4_xl). And it is much more capable than providers that claim full precision. I have no idea what they do to mess it up, maybe they quantize the kv?

Either way the provider you choose matters a lot.

>> we still need a rotating chunk of steel to stabilise the grid

>No we don't. Battery systems can stabilize the grid. You need inverters that are designed for this, but it isn't hard to make an inverter that will set then grid frequency and force it correct.

I know in theory we don't, but in practice there is no such real grid anywhere in the world that would run on large percentage of solar. So it seems we do.

I'm talking about real grid as something that powers at least thousands of households. Not a compound with a dozen buildings.

Take Canary Islands for example. They are an ideal environment to setup a 90% solar and battery grid. They have a great climate and size well suited to this. But they still run thermal plants.

Why? Because no one has successfully showed a largish scale frequency regulation/stabilisation with inverters.

See a writeup about the Texas outage a year or two ago to understand how fragile grids with too large percentage of semiconductor sources are compared with rotating masses.

You laugh, but that is exactly what I did.

I was migrating cicd from Jenkins to gh workflows for some 70-80 apps each in its own repository and support team in a random timezone. I had to open 2-3 PRs in each repo and a day or two before someone enabled "copilot reviews" in our github org.

Not only did copilot eat up all my bloody llm credits (thankfully it was almost end of the month) it created few silly comments on each of those PRs I had to respond to.

So I used an llm to write a script to respond to each of those and close them.

Also the same script wrote emails to all the teams asking for approvals, sent reminders to those that ignored it over next few days and finally merged everything in the right order.

Overall I'm a big fan of properly used LLMs.

Because it is normal these days for people to be given work they have no clue about. Then I get PRs that want to make me cry from them.

So we setup llm based reviews that catch 99% of simple problems saving us 7 out of 8 rounds of PR reviews.

Seriously. Even as simple rules as: - if you submit a PR with new X, put a link to a successful run of said X to proove it's working. - if you choose not to reuse Y, but create a new one explain why.

Those are 100% ignored by people (or their crappy LLMs more likely - so I have to teach them how to prompt, or tell them which models to use for what).

Instead of loosing sanity responding for the 10th time to someone in time zone 12h away that he tries to merge reimplementation of what we already have with no description and no test run I have an automatic llm review that tells them that exact thing 2min after they create the PR.

Same here.LLMs are really good at code review providing they are configured correctly and have right docs and prompts.

Github copilot review is s great example of the worst possible way this can be misused. It is the worst waste of tokens there is.

Instead we use a process we created ourselves that runs opus or sonet (depending on complexity) to do a review (taking into account pr description and comments - copilot ignored those) and then to re-review the issues found to cut out BS.

This has been working very well mainly thanks to the prompts and docs we made it read, as well as us having lots of example code that shows how things should and should not be done.

One crappy thing Github does is that it doesn't allow our workflow based review's tokens to be accounted to individual PR authors like "copilot review" does, but accounts tokens to the organisation. It is obvious they are trying to herd people to using gh copilot review instead. But the quality difference is like night and day. So we are staying.

Solar is great, but we still need a rotating chunk of steel to stabilise the grid. In my country (Poland) on a good sunny day half of the grid is powered by renewables (in reality almost all is solar).

But the beauty of solar is that it makes you independent from the grid. China makes the whole system. Cheap, but good quality solar panels. Cheap inverters - these are actually pretty good for the price. And finally cheap LFP batteries.

When amount of solar capacity exceeded half of entire grid, my country has established few silly rules to stop the growth.

For example let's say your house's grid connection is 24kW 3phase. If you want to connect your solar to the grid and feed in power you are limited to 24kW peak solar panels, and.... 24kWh of battery.

This is beyond stupid, but there is more... If your inverter is 10kW or more you have to give the grid operator full control over your system. They then decide when you are sending power to the grid, when you use your battery and when you use the grid. Of course they will make you push to the grid when prices are the lowest and you'll consume from the grid when they are highest.

So many people do what I did. Build off grid systems with ability to swap between the grid and the system.

Such systems do not have to be registered anywhere and I bet they are responsible for huge amount of power use.

For example before I installed solar I wouldn't even think about installing a heat pump for both air con and heating. But having extra power allows me to do it economically.