Hacker News: dandaka

New comment by dandaka in "VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO"

dandaka — Tue, 23 Jun 2026 14:24:52 +0000

Imagine you only know how to cook (use fry pan skill) and know how to cook omelette (recipe). You get the task to cook doner kebab. How many Wikipedia pages do you need to read to get a good understanding? I guess its max 5.

I think grounding your abstract problem to an example makes it more trivial, than it sounds in general.

> How would it know about Wikipedia and when to use it?

2 general concepts "You have to get good understanding of subject area before you do actions" + "Wikipedia is a good source of knowledge of subject areas" will get a model there.

> spawning a baby human, have it spend an (instant) life learning

Humans spend 99% of their life on boring repeating tasks, not learning anything, just navigating on heuristics.

New comment by dandaka in "VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO"

dandaka — Tue, 23 Jun 2026 13:35:24 +0000

Model can use tools to get that knowledge. In your example, read Wikipedia page about table tennis. Imagine a reasoning engine with a big enough context, that knows nothing. A path built from first principles to understand "table tennis spin" — does not look very long for me.

New comment by dandaka in "VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO"

dandaka — Tue, 23 Jun 2026 13:32:23 +0000

Take mathematics as an example. Humanity has found math notation, which allowed to express math rules — distill them to the core. Before math was expressed in prose — a very inefficient way, very similar to current LLMs.

In my school, math teacher was giving me prose, which I was converting to math notation. I could argue, that this prose→reasoning conversion is not required at training, and can be obtained at inference time with search tools.

New comment by dandaka in "VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO"

dandaka — Tue, 23 Jun 2026 09:09:38 +0000

I think this is a well known concept, which we can't deliver yet. LLM/transformer give us reasoning engine as a byproduct of its design, but it is quite ineffective. If we can distill reasoning, if reasoning can be achieved without general knowledge, it will be a very effective machine.

Some amount of knowledge is required for reasoning. Maybe such model can dynamically knowledge domains to have taxonomy. For example, model can't effective reason about development task, if it has no knowledge about development best practices. But population of New York or recipies can definitely be loaded run time with tools.

New comment by dandaka in "Humanity isn't ready for the coming intelligence explosion"

dandaka — Tue, 16 Jun 2026 15:14:51 +0000

"Free AI" will not fight existing knowledge or strong opinions. It will spread to empty spaces. Example with blood test — there is no podcaster there and will not be.

New comment by dandaka in "Humanity isn't ready for the coming intelligence explosion"

dandaka — Tue, 16 Jun 2026 09:03:32 +0000

AI will not help by improving extremely smart people. AI will help dumb people and dumb processes with "free" expert-level intelligence. Anyone could still ignore intelligence and make ignorant decisions. But the default mode would be highly intelligent informed decision.

Example with health - a patient can read blood test results with Opus and get very good results for "free". This is far away from helping extremely smart people, still it improves society from the ground up.

New comment by dandaka in "Waymo Premier"

dandaka — Thu, 11 Jun 2026 22:13:13 +0000

How fast is it going to expand to other markets? Asking for a friend obviously

New comment by dandaka in "Squillions: How money laundering won"

dandaka — Tue, 02 Jun 2026 18:56:26 +0000

Going from crypto to fiat and back is an extremely monitored and regulated route. It might be an easy way to settle between counterparties, but a difficult one to launder.

New comment by dandaka in "Spotify will start reserving concert tickets for fans"

dandaka — Fri, 22 May 2026 08:37:21 +0000

I imagine the best distribution model like this. We split all tickets in 2 buckets 50/50:

1/ Sorted. Some buyers have priority. They can be sorted by price paid, by amount of minutes listened, depends on the sale.

2/ Random with KYC. Everyone has the same chance to purchase.

New comment by dandaka in "Shunning AI is the human choice"

dandaka — Thu, 21 May 2026 18:48:05 +0000

You are cooking them wrong. You absolutely must ask the model to do grounding work — search online, search in your files, cross-check with different agents. They are universal reasoning engines, not a fact recalling tool.

Check my 'deep research' skill, you will get the idea

https://github.com/dandaka/skills/blob/main/deep-research/SK...

New comment by dandaka in "Shunning AI is the human choice"

dandaka — Thu, 21 May 2026 14:32:01 +0000

It is called 'jagged intelligence'. A lot progress was made in the last 2 years. Most notably reasoning models, tools use, harness progress. It takes time to build the skill to make those models useful, but they do provide a lot of value.

New comment by dandaka in "How fast is N tokens per second really?"

dandaka — Thu, 21 May 2026 14:02:23 +0000

In my experience, current agentic workflows are so slow, that for many cases it only makes sense to run them in parallel. So a lot of context switching. If we could have 10-100× faster token generation, we could have task delivery at the speed of human review.

New comment by dandaka in "An OpenAI model has disproved a central conjecture in discrete geometry"

dandaka — Thu, 21 May 2026 12:28:02 +0000

Not sure I follow your point.

The fact that humankind grew from 5M to 8.3B, while dramatically improving longevity and quality of life speaks volumes. Multiply life quality × population × life duration, not only "misery and destruction" is not the case, but you could rather see powers of positive technology influence.

New comment by dandaka in "An OpenAI model has disproved a central conjecture in discrete geometry"

dandaka — Thu, 21 May 2026 09:57:10 +0000

Do we have a list of open problems? Would love to see a chart, where AI solves such problems one by one in the upcoming years.

New comment by dandaka in "An OpenAI model has disproved a central conjecture in discrete geometry"

dandaka — Thu, 21 May 2026 09:36:50 +0000

10000 years ago we had 10-15% deaths from violence (skeletal evidence). As well as infections, child mortality, starvation and injuries.

Benefits of civilization eliminated most of that + increased quality of life dramatically.

New comment by dandaka in "An OpenAI model has disproved a central conjecture in discrete geometry"

dandaka — Thu, 21 May 2026 09:31:38 +0000

We can reach a different situation

1/ No one knows how even small components work, because their inner working mechanism is too hard to understand by human mind

2/ The whole society is run (in intelligence sense) by alien minds

New comment by dandaka in "Reimagining the mouse pointer for the AI era"

dandaka — Wed, 13 May 2026 09:38:03 +0000

The goal is not to record activities of others, it is to have full context of operator activity. More like a personal knowledge base, that camera pointing to a cop.

New comment by dandaka in "Reimagining the mouse pointer for the AI era"

dandaka — Wed, 13 May 2026 09:36:45 +0000

why not? if context stays on device and operator is in full control, what downsides are there? there is no observer here, only operator and full context of his activity at his fingers

New comment by dandaka in "Reimagining the mouse pointer for the AI era"

dandaka — Tue, 12 May 2026 19:40:09 +0000

Next generation of OS should have constant video and audio recognition by on device LLM. This will provide valuable context for a lot of scenarios. So instead of frequent copy-pasting we are used to, we can let agents access context of our whole workflows from different apps.

But Google is a very ill positioned candidate for such OS. I would rather trust Apple and local-first on-device models.

New comment by dandaka in "AlphaEvolve: Gemini-powered coding agent scaling impact across fields"

dandaka — Thu, 07 May 2026 16:47:50 +0000

How many times we have to hear again about Erdös problems? :) It sounds like a great achievement for humanity at first, but after a while they keep coming back!