Agreed, alignment is an inside joke. The fact that they admit that it's done by... AI is quite revealing. About the job losses however, either that tech isn't as useful as it's hyped to be, or it's so useful that it's creating new jobs. Just saw this: https://www.newyorker.com/news/the-financial-page/has-the-ai...
I'm still here, nearly 50 years and counting. If you had asked me what I imagined AGI would look like back in the 90's, I would have told you "A system that can do everything we can: see, hear, think, do.". If you had shown me GPT-6 back then, I would have said "It looks like a really powerful program, but that's not really what I had in mind.". That's AI, but it's not quite general.
And then you'd ask it about an area you're knowledgeable in and realise it routinely makes stupid mistakes.
Or you'd ask it to add a new page to your website and shout at it to use your existing brand colours instead of inventing some and realise it's not AGI at all...
Origami design will be my personal test bed for the coming years.
It's objectively very difficult and technical, it's spatiovisual, it's artistic, learning resources for it are sparse and most just learn by the FAFO method, current AI sucks terribly at it, and it's not likely to ever be specifically targeted by benchmaxxers.
Scientifically useful physics simulations. Every model absolutely sucks at them.
Or, as someone else points out in another thread here, academic writing. It's one of the things newer models seem to have actually gotten worse at. Even when you give them detailed instructions on how to write and what to avoid, the "load-bearing", "A but not B" and journal-like writing make it in anyway, with the supposed AGI having no ability to reflect on how blatantly unacademic (and often unreadable) its writing is.
Likewise, it's very impressive and useful, but it is obviously not AGI to those of us from that era.
If anything, the fact that it is so powerful is almost a concern, because I think we are still way underestimating what these systems will be able to do when we give them more cognitive capabilities.
At the moment we are something like, having had great success with propellers and have promised we will fly to the stars.
People love to say 'this is the worse they will ever be', then extrapolate to conclusion that they will continue to accelerate at the same rate of progress of last few years .. it may, maybe, or we will hit a ceiling, might be a temporary one, could be 5 years or 50 years ..
Anyone that is not impressed by what ChatGPT or the likes are doing now is being either dishonest or is incapable of being impressed.
Only the translation and language understanding capabilities are enough to be impressed, and they are 2 year old already. Now, the AI do see, draw, speak, listen, think, work, etc.
Someone from the 90's would simply not believe that the AI would be a machine but would think for sure that a human is behind. The only odd thing would be that this human would both exhibit high intelligence and stupidity at the same time.
Oh God, that "a dead drag handle during a booked half-day you do not get back" got me. I saw this pattern in Claude's 'explanations' so many times. It's trying to say that it did something significant, and that you'd only have found out much later, at higher cost (or something). That annoys me to no end.
I use Claude Code (both Desktop and CLI) all day every day (as well as OpenCode with DeepSeek V4 among others). For non-mission-critical stuff (small internal tools, anything that doesn't touch the client and doesn't handle critical data or infra), I'm on Auto (sandboxed in the working directory). For everything else I approve every change, meaning I review every code change the model proposes. Plan mode is great to write the architecture as well, as it doesn't touch any file, doesn't write anything, and lets me iterate until the arch is perfect and the proposed high-level code (modules etc) is solid. I haven't used an editor in over 4 years, apart from writing specs.
How do you review the code? I can use git to see changes, but it's super annoying to tell the agent file line plus comment for each change, so I catch less stuff early which seems to compound.
Ornith 1.5 35B MoE is very close in most benchmarks, but runs at 100 tk/s where Qwen3.8 27B runs at 30 tk/s (MBP M5 Max 48GB). It's now my go-to model for quick inquiries. Ornith 1.0 had problems with world knowledge, but 1.5 seems very solid on that front. I'll be running some agentic tests today.
Ornith 1.5 35B is so fast that my CPU doesn't even have time to heat up, where Qwen3.8 would turn my laptop into a toaster within a few minutes. I hope agentic tasks will hold up!
I ended up getting an M5 Max MBP with 48GB and it runs Qwen3.8 in Q4. I developed 3 apps so far, medium complexity, with no issue using OpenCode. This is probably the minimum setup for comfortable local agentic coding in my book.
They haven't communicated well the fact that the default model should be Sonnet 5, which should give you unlimited use for common coding tasks (say with occasional subagents use) on the Pro plan. Instead they're pushing Opus and even Fable, to try and get people addicted to the higher tier, without realizing that nearly everybody has a Sonnet for peanuts via DeepSeek V4 on OpenRouter, or completely free through Qwen 3.8 locally. I predict major trouble for Anthropic, now that OpenAI's models are closing in, are cheaper for daily use and don't have those silly 5 hours limits - and that Chinese models are getting really good and are even cheaper.
Sonnet 5 is a trash model and stupidly expensive if you accidentally set reasoning tokens high - more expensive than fable - it absolutely should not be the default lmao. If the common coding tasks you use ai for is doable with sonnet or local qwen - you're either not using Claude code (if you are, you'll very quickly see that sonnet 5 in Claude code is not a model for "occasional subagent use" - spinning up subagents is the only thing it's good at and it does it way too much. It can spin up subagents and waste huge amounts of tokens but it can't write good code lol.) or you've got Claude code workflow that is very human in the loop where you are significantly steering and controlling the models. And in that case your default should be to use gpt. Claude models are stupid slow.
Claude Code also doesn't make it easy to make _efficient_ use of the different models to reduce overall cost. There are many tokenmaxxing features (e.g. ultracode that spawns dozens of subagents) to burn through the 5-hour limit in minutes, but if you want to let an Opus planning agent use Sonnet for implementing you have to orchestrate your own workflow. I'm pretty sure that's because the Anthropic employees working on Claude Code have unlimited token budgets so they're mostly on tokenmaxxing workflows themselves.
Qwen 3.8 is excellent. With the right harness, it does about 95% of what Opus can do, in my case automation software development. Since 3.8 came out, I have significantly revised my expectations for a local model. Give it another year or two, and we'll be running fast and free local models for nearly everything that matters, and these costly subscriptions will be a thing of the past. I've always believed that AI should be free for everybody, like TV and radio. We're almost there.
Free tv and radio? Where do you live? Where I live you either pay taxes for it, alternatively, it is so ad infested that it is not possible to watch it.
I suspect the same will/is happening with AI. Either you will pay for it, or it will be so ad infested that it will become useless.
Not op, but check out open code; you can turn on K/V quantization to help with increasing context if you have not already. I think K needs to stay at least 8 but I hear V can go down to 4?
I am running it on 32GB and I did not saw model loosing it context even after 4-5 compactions in pi. I am running sessions for few days sometimes. I think it looped once, but loop police extension stopped it. The only problem I have know is how pi compaction works, which is forcing full prefill which takes time and it is erroring a lot. I wrote my own compaction that should remove full prefil but it does not work. But this is the only problem with this setup and it is more problem with pi then the model. I much more prefer it to use Qwen then paid models: Claude forces me to do reauth every other day and codex models either are too costly or not capable enough.
I've a 32gb and 64gb (work) MBP. 32 works - just and sits at around 28/29gb of 32. 64 works great, so the 48gb laptop with MLX + MTP should be fine. I'm using Ollama.
I initially used the Claude Code harness on 3.6 A3B, but found that tooling would break as Claude released new versions and things would go weird. I've since written my own harness which has basic operations: read, find, bash (which can write files, python etc...) & web_fetch, all within a mac container. Works amazing. You don't need anything complicated to go very far.
Low hanging fruit would be Pi or OpenCode. If you really want a much better understanding of what your hardware is capable of then give writing your own a go.
Additional tip: Low Power mode reduces some token speed, but stops the laptop over heating and the fans going crazy.
reply