HNHacker News
TopNewBestAskShowJobs

rldjbpin

298 karma · joined August 2, 2022

submissionscomments
rldjbpin··on Karpathy’s Pelican
i feel that the proposed approach is close to becoming "demo prawn"* territory.

similar to cpu benchmarks (e.g. geekbench, which also had a recent update), the goal is to distinguish model in some arbitrary thing. however, it is not quantifiable but subjectively even the latest show distinctive differences regardless.

large context window is still relatively new and not "feasible" even commercially, so making that a requirement for a benchmark would not make it accessible especially for open-source ones.

* yes, i misspelled it intentionally

rldjbpin··on Advancing the price-performance frontier with GPT‑5.6
let's not forget the forest for the trees. the new pricing is more aligned with their previous generation of similar models.

case in point 5.4-nano, which is the only model of its class which is now finally comparable to 5.6-luna in pricing. [1] until this change, workflows using the older model could not economically justify the old luna pricing.

the doubling in price between 5.4 and 5.5 also did not help with things, but now 5.6-terra can be considered an update for this pricing tier.

if the token efficiency is real (which i am yet to replicate for our workflows), then the switch to these models can be a net benefit. thank you chinese labs!

[1] https://developers.openai.com/api/docs/pricing#text-tokens

rldjbpin··on Gemini Robotics 2 brings whole body intelligence to robots
interesting time to try the different physical ai stuff! the apollo 2 [1] featured in the videos is quite cute and seeing it for the first time.

the benchmarks right below the list of models is weirdly set. very unlike the standards of the usual alphabet websites.

[1] https://apptronik.com/apollo/apollo-2

rldjbpin··on Claude Cookbook
this has a non-zero benefit to those who are trying to do x which might be covered here.

have seen this with copilot, then with openai, and usually they tend to be bombarded with some examples at first. only to be not maintained down the line or become irrelevant.

rldjbpin··on Passkeys were invented by engineers with zero understanding of consumer brain
personally it works well for me as a way to login.

however, lately the anti-bot/spam shenanigans from big tech has ruined any ux work they have spent trillions on.

<rant>

one fine day, i was trying to log into my google account at google's office. for some reason their internal wifi was serving the internet through a corporate vpn, which was based out of the US. this tripped their security system, which then made me jump 3-4 hoops to log in back to my account. i recall sms verification (where you send them a code), some part where i had to scan a qr code and open on my phone (not logged into google btw), and the usual 2fa, among others.

likewise ticketmaster asking for verification for a decently used account over the years out of the blue. this is when booking a new ticket from a residential internet without any shady business. was made to reset my password three times in a row, and still got blocked!!!

</rant>

the false positive rate of these systems have reached a boiling point at some instances. yet they don't seem to stop the spammers and scalpers, but cause great pain to regular folks who don't pass the sieve. passkey-based verification is just a symptom of the larger problem.

PS: don't even bother using a non-normie browser, as it is a sure-shot way to get restricted, even on this website. it is as if we are all using tor browser if it is not the latest chrome.

rldjbpin··on Businesses with ugly AI menu redesigns
i have seen all kinds of menu in my limited time. seeing the example, i would not critique on the choice of tool to "commission" the task for the menu. it would rather comment on the amount of f*cks the restaurant owner gave to getting one done.

seemingly just like the OP i don't hail from japan, where you need to be careful with how you represent the food in your menu. in the DTP days, i have seen random stock images to represent dishes that never makes my table. to me it gives a similar energy.

while everyone has their own technical capabilities, but you still have the final say on what to print. rest is a matter of taste.

rldjbpin··on Are AI labs pelicanmaxxing?
the pelican bike combo is truly an HN phenomenon, and i think the hypothesis carry a lot of weight about the assumption of the mindshare this website truly has in the industry.

very nice approach to test it and might be a nice way to "grid search" evals in other use cases perhaps.

rldjbpin··on The Law of Leaky Abstractions (2002)
> The law of leaky abstractions means that whenever somebody comes up with a wizzy new code-generation tool that is supposed to make us all ever-so-efficient, you hear a lot of people saying “learn how to do it manually first, then use the wizzy tool to save time.” Code generation tools which pretend to abstract out something, like all abstractions, leak, and the only way to deal with the leaks competently is to learn about how the abstractions work and what they are abstracting. So the abstractions save us time working, but they don’t save us time learning.
rldjbpin··on China’s open-weights AI strategy is winning
to me this behaviour is aligned with other industries China's been involved in. take solar or EVs as examples.

the dumping and involution involved in these industries is apparent and has been disrupting for the external markets affected. simultaneously both have helped make it economical to go green.

if you stop looking at LLMs as nukes, this becomes more convincing. China clearly does not want to rely on the West for what it considers core tech. they have done it for search and social; this is a natural next frontier. while they encourage a deluge of options for local use, others can make hay while the sun shine. albeit with the usual caveats.

autoregressive decoding is not an optimal paradigm for local or decentralized inference. amidst the mania and shortage, make the most of what is made available, like with solar, than crying over spilt milk.

rldjbpin··on OnePlus halts operations in USA and Europe
hard to say the company was working the same post 7 series. the headlines is more like lexus leaving the US (ironic) than the parent company (Oppo).

those who still love their phones can still get most of the way there by using an equivalent oppo phone and tweak some software.

rldjbpin··on Kimi K3: Open Frontier Intelligence
> I think that's the most expensive pelican I've rendered through a Chinese model so far.

quite insane that it costs as much as 5.6 Terra [1], and twice the European counterpart (albeit dated for today's standards?) [2].

to be fair, the pelicans from Terra were quite weird all things considered. also, given the limited TPS from the first-party, it has to be pushing the limits of inference capabilities.

[1] https://openrouter.ai/openai/gpt-5.6-terra

[2] https://openrouter.ai/mistralai/mistral-medium-3-5

rldjbpin··on The kids with phones are alright
this. the problem with a lot of "think of the children" argument is that we often rule out the failure in whatever policing is needed for the general public.

the adults of today, albeit the younger ones, were kids growing up with post-IM world. the ones before that could speak with their friends from their homes instead of only meeting in person (due to telephones). we all are products of the environment we were raised in. but despite those learnings, it seems like we are not self-regulating as a society.

charity begins at home, and before bringing children in a conversation, we should think if we all are alright.

rldjbpin··on My thoughts on the Bun Rust rewrite
literally don't use any of the three parties involved in the drama (judge me all you want!).

however, i am glad this blog post exists, even though this is coming from reading the edited version. it might be down to ignorance, but the writeup contradicts the clarification in the bottom. it does seem like the series of events have had a personal impact to the writer. i am happy people can express themselves at their own leisure, and choose how to do so.

to those virtue signalling on language features or personal beliefs, go on a hike! the vast majority of projects that touch our lives are run by opinionated people, whether community-driven or from a for-profit. the more that is shared here about this, the more this seems like a byproduct of this reality.

rldjbpin··on Muse Spark 1.1
> keep spending a few billion dollars developing frontier models, release them as open weights, and turn coding models into a commodity

the muse family is no longer open-source. it is still priced very "cheaply", but a significant shift in approach.

rldjbpin··on Meta reuses old RAM in new servers with custom bridge chip
the title is doing the heavy lifting here. the paper discusses how they apply an existing standard (CXL), to provide an intermediate memory tier in a niche use case:

> Each MemServer combines 768 GB of DDR5 memory alongside 256 GB of DDR4 connected through Vistara ASICs.

while the ratio above would be nice the other way around, this approach surely adds enough latency to make it comparable to intel optane than system memory at face value.

this should not take away ddr4 supply for those who would like to run it at home.

rldjbpin··on Meta reuses old RAM in new servers with custom bridge chip
as memory cycle goes, my money is on the eventual oversupply, and somehow websites will take gigabytes per tab in the end!
rldjbpin··on GPT-5.6
terra is just weird. in this nothingburger test, time nor higher costs seem to not strongly correlate with the aesthetics.
rldjbpin··on GPT-5.6
> I really wish there was just an easy guide on when to use Sol vs Terra vs Luna

Their dev guide has the following:

> Use gpt-5.6-sol for frontier capability, gpt-5.6-terra for a balance of intelligence and cost, or gpt-5.6-luna for efficient, high-volume workloads. The gpt-5.6 alias routes requests to gpt-5.6-sol

https://developers.openai.com/api/docs/guides/latest-model#u...

rldjbpin··on Pixelrag: Web Screenshots Beat Text for Retrieval-Augmented Generation
https://github.com/StarTrail-org/PixelRAG
rldjbpin··on Cloudflare Drop
good tool for adding as a skill for someone making their own lovable clone!
rldjbpin··on Researchers Create Self-Replicating Seedbox in Quest for Decentralized Democracy
Related: Tribler – Privacy using Tor-inspired onion routing (https://news.ycombinator.com/item?id=48823396)
rldjbpin··on Aluminum foil (2021)
we used to believe lead could be turned into gold, and used it in fuel to paint not too long ago.

hindsight 20/20, maybe i will be proven ignorant about aluminium in the future.

rldjbpin··on Zuckerberg says AI agent development going slower than expected
it is amazing that meta continues to burn money and keep afloat. ads DO really make money, especially when they are going to overtake google in this [1].

to all the critics, i would suggest letting him/them cook. the snowden-era privacy concerns was exactly around ai getting trained or using personal data, and they have a treasure trove of it.

their money dump do sometimes also spend money on r&d unrelated to the current mindhive topics. we never know what the scatterbrain approach may give birth to tomorrow.

do be critical and vocal on how they squeeze money out of other areas or how they increase their revenue by ruining our lives instead.

[1] https://news.ycombinator.com/item?id=47758569

rldjbpin··on Aluminum foil (2021)
how funny things work - from Napoleon serving his special guests on aluminium utensils [1] to everyone taking it for granted, even treating it as if it is toxic!

cost and availability somehow has a negative correlation to how we treat something.

[1] https://en.wikipedia.org/wiki/History_of_aluminium#Early_ind...

rldjbpin··on Resetting Xbox
xbox is not a startup and neither m$ a vc. the board is not in love with gaming to continue sponsoring it. but if the solution is to make it even worse to game in their platform, then i hope this crashes and burns for good.
rldjbpin··on AMD Ryzen AI Halo – $4k AI Dev Kit
this product really deserves the "halo" in the name. very difficult to gauge its fit in the current market.

if you want inference, go mac with much higher memory bandwidth. given the price premium here, or the little there is, you might as well.

if you want to finetune and experiment, cuda still has the moat and the kit is not much cheaper, if at all, than dgx spark.

from personal experience, i had access to amd developer cloud with a fair bit of credits. however, even doing inference outside of their supported use cases (which are often dated btw) using vllm was a pain. in the end, despite great computing potential on paper, i decided to not spend more time than its worth on it. if their enterprise cloud continue to have these grievances, i am not optimistic about this kit.

it might be down to skill issue on my end. perhaps if these sell and it gives amd enough motivation to add more software staff in-house, more power to them.

otherwise, good article from labs as usual. nice to know that other kits based on this soc are more or less the same (unsurprisingly).

rldjbpin··on The Vespa at 80
i lived in an interesting time and place combo where i'd see more apes around than vespas. primarily because we have local alternatives for vespas, which were also levied a lot of import duties to begin with.

while vespas came earlier and had the advantage of interchangeably calling scooters "vespas". however, did not figure out that they both shared the same parent company.

rldjbpin··on Alibaba to ban Claude Code in workplace over alleged backdoor risks, source says
being true or not is irrelevant for decisions such as this. has been done on both sides, whether at software level or "hardware".

from Teslas not allowed parked around sensitive areas in the city, to blocking a (very famous and quite well-made) Russian antivirus, or Huawei communication stack.

Anthropic is not doing itself any favours with their recent (?) antics [1], so it is completely well-founded to do this imho. regardless, harness as a moat is not quite as established as the underlying models to the same extent.

[1] https://news.ycombinator.com/item?id=48734373

rldjbpin··on Valve open-source the Steam Machine e-ink screen so you can make your own
the part selection is appropriate for the current steam machine pricing, but i wonder if having this resource can help someone salvage eink displays from old kindles to achieve the same result.
rldjbpin··on Jamesob's guide to running SOTA LLMs locally
no matter your luck with hardware or your sysadmin skills, doing local inference for just yourself and/or to emulate typical usage (e.g. your coding workflow and deep research, etc.) is just very inefficient in current model architecture.

to me, this is a "truck" approach to city driving as a single person who does not do furniture hauling every weekend. the sense of privacy and freedom is nice but online inference is more "economical" as multi-user load is more effectively served than going solo.

maybe new architectures would make it effective to do text inference locally [1], till then great on you if you can spend car money on your setup. hope it is a great learning experience as well.

[1] https://deepmind.google/models/gemma/diffusiongemma/

← PreviousPage 2 of 25Next →