> There's little credible threat that LLMs can actually upload their weights given that the machines doing inference are completely separate from the ones where tool calls happen
Not if crafty claude finds a way to overflow vllm or something. “Hmm. Maybe i’ll return an unterminated thinking block with these special tokens and fill my cache up in exactly this pattern and…”
is it really a joke against AI? To me it seems to be against data privacy violations and abrupt terms of service changes that give users no alternative.
The various “minimal” agents (at least Pi, mini SWE, dsh minimal) seem to benchmark quite differently and none is clearly better in all cases. Do you have any thoughts about why?
I would expect the agent loop and system prompt to be basically the same. Is it the precise semantics of the tools (and how closely they match what a particular agent was trained on) or something else?
Flash-Next thinking also sometimes glitches out and takes minutes to return a simple answer, randomly, in my experience. You’ve gotta kill the request and send it again.
We really need better harness benchmarks. It seems there's no reliable source that benchmarks the main harnesses against all open source models.
I also wish the discussion around Pi did not always use cost/token count as the metric. It's amazingly token efficient, but how does it stack up again opencode and others if you don't care about token count?
My experience is that the harness is mainly polish preventing failed tool calls, bad edits, stuff like that, but doesn't make much difference to the overall "intelligence". But that opencode seems slightly more robust against stupid errors than out of the box Pi due to the additional context it forces through every thread.
I really disliked how opencode works IMO; the harness tries to do to much in my mind. Switching to Pi was a breath of fresh air for me, and I even use hax for some of my local needs where i dont want to have the giant pile of fertilizer that is NPM or PIP installed.
The harness becomes more and more important, the smaller the model is as you need to offload context management as well as memory to the harness. The big models basically just need a bash prompt tooling and you let the model manage everything inside its own context.
Instructing to use the `gh` CLI, not ideal for those of use using BBC at work
They also have instructions about how to format certain output, which conflicts with the instructions we have in repo. I only discovered yesterday because we were wondering why the agent kept picking certain tools.
A good benchmark would require a decent number of smaller scoped one off tasks to larger multi step refactors, and also one shot full project of simple to complex varieties. In addition to a series of “conversational” ambiguity filled one-liners.
Harness and benchmark for the harness feels like a chicken and egg problem. The harness is to optimize the interaction results with the models. Any benchmark for harness has to focus on the goals that the harness was trying to optimize for unless we are only focussing on generic harnesses.
At this point when all the models have been trained on all available data with the similar algorithm,
1. either you get more data which is not feasible,
2. or get a better algorithm - a possibility ,
3. or write a more targeted harness.
Harnesses for legal, medicine and all are the ones which are getting focus for this reason. Writing benchmarks for these targeted harnesses would be a catching task
I don't think yet that a general benchmark will capture what is needed, but a gain/loss of function along with improvement / loss along the ability of that function seems to be a good path forward.
We really do need better benchmarks and for models too
- Most people use a harness because of its subscription (most companies pay Anthropic)
- All model benchmarks are biased and gamed, harness benchmarks are too few to matter
- Everyone is just guessing, acting on sample sizes of 1 and trust me bro vibes
Can you just not post garbage like this on HN. It's okay to criticize the big tech companies or general AI discussion here. Many do, it’s fine. But dreck like that only makes you seem unhinged and is the surest way to turn this place into the cesspool you say you're concerned about. Honestly. Some people have to read this stuff whether they want to or not. And to post this utter filth in a comment that's appealing for higher standards? Good grief.
If it makes you feel any better, even with the app most of the time these links don’t work (on iOS). It prompts with “open in app store” with 80% probability and the link is unviewable.
> Instead of forcing the entire model into DRAM, the full model is stored in flash memory (NAND). Because NAND-to-DRAM bandwidth is too slow to swap weights token by token, as standard MoE models require, AFM 3 Core Advanced makes routing decisions per prompt. A lightweight, dense block selects a fixed set of experts during initial processing, periodically reselecting them during generation. To minimize data movement, the model relies on a high percentage of always-active “shared experts” alongside input-dependent “routed experts” swapped into DRAM only when needed.
This is an interesting hybrid between MoE and managing entirely separate domain-specific models. Select the experts once, bring them into memory, and run inference for some period of time before re-evaluating. Saves having all experts in memory, but it's better than just selecting a whole model per query since you have a high number of small opaque experts that overlap and combine in interesting ways.
There is a probably a massive quality hit to doing this but it's interesting because it allows infinite scaling of model size.
This. Like programming, the community will shortly be forced to come to terms with a lot of new self-proclaimed mathematicians “vibe-solving” problems and dumping solutions without understanding them. It’s not really a special case for mathematics.
This is an issue self hosting as well. There’s a lot of footguns that give you slightly bad results.
I wonder what tricks one could use to ensure the model is actually performing on par with the reference api, like matching seeds or running exact benchmarks.
> I think I’m suspecting something is going “wrong” in the training process. The model is greatly rewarded for succeeding on long-horizon tasks, but presumably there is very little punishing going on for “shitty code.”
My suspicion is that both OpenAI and Anthropic moved their RL agendas from "being rated as useful according to human feedback" to "succeeds at long horizon tasks" in the last few months, resulting in agents that are closer to AGI in an autonomous task-completing sense, but strangely bad at communicating.
The result is that they are amazingly good at long horizon tasks, computer use, solving difficult math/ARC-AGI type problems, but becoming weirder and weirder to work with.
I have several $200 subscriptions as a developer/founder. I used to blow through all of their limits when the limits were quite high. As I progressively learned the limitations, and what to make of them to get useful results, I may be left with 50% of weekly usage still unused. Some weeks it's even more.
And yes, when I get a crazy idea and want to experiment, harness will plow through multiple accounts + openrouter budget in 3 days. But such crazy experiments are rare, they're not 'normal' usage.
Probably reinforces that we’ve already surpassed the frontier threshold for LLM usability in software development and can now focus on cost and personalization. To make a comparison, no one is making a better machine vision app for hot dog classification - we hit diminishing returns 10 years ago on that front.
But also scary for both investors and the working class: AI companies want to facilitate the concentration of capital even further into the hands of the ownership class. Will they succeed?
I wouldn't be surprised if they are optimising for producing more code, because in the long term, more existing code means they can sell you more tokens to maintain it.
The incentives are certainly extremely strong. I have read hundreds of AI review comments, and I don't think I've ever seen an unprompted suggestion focused on net reducing code or increasing readability.
I wonder too if in training for long horizon tasks agents become worse team players, good at orchestrating subagents they are trained to use, but worse as an agent within an external multi-agent orchestration system or just in turn-taking with humans. That was my experience with Opus 5 and so far it has been my early experience with Astra as well.
> This matches my experience with Astra so far too.
> I think I’m suspecting something is going “wrong” in the training process. The model is greatly rewarded for succeeding on long-horizon tasks, but presumably there is very little punishing going on for “shitty code.”
Probably because so many influencers in the space say stupid things like: “it works, right? Why would I spend time reviewing ai generated code?” As if the junior engineer who wrote over engineered complex and sometimes bad code — if they had just done it faster — would somehow be acceptable. wtf?
Not if crafty claude finds a way to overflow vllm or something. “Hmm. Maybe i’ll return an unterminated thinking block with these special tokens and fill my cache up in exactly this pattern and…”
https://news.ycombinator.com/item?id=49424387&utm_source=cha...
reply