Philip Nuzhnyi

5 Jan

“Men behaving like machines paved the way for machines without men.” - David Noble

Good overview of NVIDIA's RTX Spark chip - app.tldl.pro/episodes/ant...

"Local compute that keeps your data secure on device is the best case for these AI models. NVIDIA, ironically, the company that's chiefly pushing the AI data center build-out worldwide, may end up showing us a better way forward"

Good overview of NVIDIA's RTX Spark chip - app.tldl.pro/episodes/ant...

This thing here www.nvidia.com/en-eu/produc...

Failing forward. Lesssgo

Gordon@gordon.bsky.social

"Error is, in itself, neither bad nor good, but endemic - it cannot be eliminated. In fact careful consideration shows it is error that drives the system!" - Glanville, 2019, "Try Again. Fail Again. Fail Better.”

Oh wow

threearrows3.bsky.social@threearrows3.bsky.social

Friedrich had a very distinctive style.

beaconbits.appFabrica Moderna Caparica | Beacon BitsBeacon dropped by @callmephilip.com

📍 Fabrica Moderna Caparica Other · Av. Gen. Humberto Delgado 19B, 2825-286 Costa de Caparica ☁️ 19°C Working #BeaconBits

Philip Nuzhnyi reposted

Evergreen Intel 👻

@vcdgf555.bsky.social

Constantine could use some help in getting non-lethal aid to Ukrainian defenders on the front. Would mean a lot if you could support this cause with a repost, like and/or donation. Good luck @teoyaomiquu.bsky.social and thank you for what you do. 💙💛

Constantine@teoyaomiquu.bsky.social

23,000$ fundraised over the the past 14 hours 50000$ left to go. Your donations, likes and reposts are extremely helpful.

Congrats to @gyroscope-app.bsky.social on their 12th birthday. There is a cool retrospect of the app's iteration here - gyrosco.pe/museum/. I'd love to see support for atproto based storage and aggregation. Your vitals on your PDS, community maintained appviews to get different perspectives on data

Lexicon garden is awesome - lexicon.garden. Shout-out to @ngerakines.me

"This paper was written with human assistance." - github.com/openai/math/...

UX design and software architecture design. Designing and maintaining software architecture both humans and machines can easily reason about is hard

entropy@entropy.obscurityskies.xyz

UX design might be the most important skill in software right now, and for the foreseeable futurable. Any SWE can spin up the most convoluted system design and have it realized for nothing. Most have no clue how to make the ergonomics of using the thing pleasant for humans (and neither does Claude)

SAGACITY.md

Pavel@spavel.bsky.social

ETHICS᠎.md

Still waiting for my jev to arrive. In the meantime Sean Goedecke has something jevonian using existing models here github.com/sgoedecke/sy...

! Do 1 round through the model (as opposed to autoregressive loop where you feed results from the previous batch into the next iteration and keep assembling final result). Only grab logits for the last chunk (what comes after "choice_index:" in our case). Important: send multiple questions in 1 go

Get rid of any results that are not our supported choice indices. Turn logits (random looking numbers like 3, 89, -386) into a list of probabilities that add up to 1. Get your answers with probabilities back.

Still waiting for my jev to arrive. In the meantime Sean Goedecke has something jevonian using existing models here github.com/sgoedecke/sy...

Figure out model's internal encoding representation for the IDs your system supports (up to 255 choices in this case). This extra PREFIX business seems to be dealing with possible different ways to encode the same value (e.g. "2") depending on whether it appears in "2" or "choice_index: 2"

Massage your prompt so it looks like "user says / assistant says" with our prefix at the very end so that the model is set up to spit out a single value for one of the choices.

Still waiting for my jev to arrive. In the meantime Sean Goedecke has something jevonian using existing models here github.com/sgoedecke/sy...

Walk through SystemOne inference github.com/sgoedecke/sy...

Each question is framed as a _numbered_ multiple choice question. E.g. Options like "blue, red, yellow" become 0, 1, 2. The model will answer with the number.

Still waiting for my jev to arrive. In the meantime Sean Goedecke has something jevonian using existing models here github.com/sgoedecke/sy...

I am no DOOM specialist but Sean's model does not seem to be super good at it yet - github.com/sgoedecke/sy.... I am old enough to remember the days when we benchmarked new models using pelicans on bicycles

What is the approach? According to Sean > As long as you have access to the logits (for structured outputs) and can prefill data into the prompt, you can turn any LLM into a general fast classifier.

17 Sept

Dmytro Apostle is doing a breakdown of AI slop videos russian soldiers have been using to report about their successful missions to their immediate commanders and beyond - www.youtube.com/watch?v=MrSn...

17 Sept

Looks like a whole lot of claims around territorial gains and control are not even staged IRL performances - they are utter AI SLOP.

17 Sept

If you find yourself lost in the woods, fuck it, build a house. “Well, I was lost, but now I live here! I have severely improved my predicament!” - Mitch Hedberg

16 Sept

Note to self and a quick reminder: A fit body, a calm mind, a house full of love…

getair.spaceairspaceThe database you already have.

15 Sept

This looks great

daniel r(ooooooh) in nyc 🇺🇸@danielroe.dev

🙋‍♂️ ever wanted to build something without needing to configure a db, auth, file storage or a cms? it turns out you already have all four if you have a bluesky account 👉 airspace is a typed client for it

15 Sept

Seems to be working directly from mobile inside an RN app after applying some elbow grease - pdsls.dev/at://did:plc...

14 Sept

Oliver Burkeman's "done list" running on atproto? www.oliverburkeman.com/donelist Anyone doing "everything tracker" here?

14 Sept

🫛(doom) 5% 🥕(boon) 56%

Philip Nuzhnyi reposted

Jake Gold

@jacob.gold

13 Sept

In case you weren't aware how the anonymous Hacker News moderators control the narrative in Silicon Valley. (Bluesky Communities can't come soon enough)

Jonathan Warden@jonathanwarden.com

Yeah that story got penalized. Orange is the front page ranking, black is the ranking it should have given the raw ranking formula with no penalties. HN moderators often apply penalties to news and political stories, even if they are very tech-adjacent. news.social-protocols.org/stats?id=496...

11 Sept

I found Jesus

Daniel van Strien@danielvanstrien.bsky.social

Need a historical illustration? Search 1.49 million images from British Library books and Britannica (1500s–1920s). Find similar images, download full-resolution images or transparent PNG cutouts, and explore the sources. huggingface.co/spaces/davan...

youtu.beChasing big models is a curse. Fix your workflows instead.YouTube video by ed3d dot net

11 Sept

Mr Harford did it again. What a banger!

Tim Harford@timharford.ft.com

Cautionary Tales - Talent, Time and Money: The Curse of General Magic timharford.com/2026/09/caut...

Philip Nuzhnyi reposted

P(aul) Frazee

@pfrazee.com

6 Sept

We should stop thinking about software as anything other than getting the computer to do what we want it to do for us

Philip Nuzhnyi reposted

@ed3d.net

3 Sept

So this is the start of a series I've been working on to explain, in a coherent way, how we should be thinking about actually using LLMs to be useful, instead of the LinkedIn-esque "I burned $5000 on Fable AND--" The thumbnails aren't going to get any less dumb, though. That's the Ed Promise.

1 Sept

Aug 31 Weeek in podcasts From "The Founders" on Claude Hopkins - he tours Schlitz's brewery to help them figure out how to advertise to get from 5th to near 1st most sold beer app.tldl.pro/episodes/430... "The maker is too close to his product. He sees in his methods only the ordinary." ....

Another one from Techbrew Ride Home on OpenAIs ad business - "$1 billion in terms of an annualized revenue run rate ... [for] its ad business, which is roughly **200 days old**" app.tldl.pro/episodes/50-...

3 Sept

Techbrew on Fable 5.1 - app.tldl.pro/episodes/hey... "[...] changing the economics of running persistent agents, reducing the cost of cached context by 75%, and introducing a new security architecture [...] designed to let organizations retain monitoring data inside infrastructure they control"

"He does not realize that the world at large might marvel at those methods and that facts which seem commonplace to him might give him vast distinction. Tell the pains that you take to excel. [...] Serve better than others, offer more help than others, and you are pretty sure to win."

2 Sept

Techbrew Ride Home on "Hugging Face rogue-AI postmortems" with an extra take from @emollick.bsky.social "agents should seek human help for approval, expertise, variance, and interest, warning against automating the wrong half of the job". - app.tldl.pro/episodes/50-...

Notes on why LLM generated content is so dense and hard to read + what can be done to improve it ⤵️

LLMs are being monotonously pompous with their writing - and it’s this monotony that smells like slop

How to fix this? One possibility: banning patterns in the system prompts. Tricky side effect - this might damage model’s reasoning. Another approach : don’t ban anything, let models write freely and de-cliché in a second pass

Notes on why LLM generated content is so dense and hard to read + what can be done to improve it ⤵️

The AI tells are actually ancient, respected techniques: - parallelism: repeating the same sentence shape for rhythm - antithesis: «it’s not X, it’s Y »; made by Aristotle, perfected by JFK -> “Ask not what your country can do for you — ask what you can do for your country” - tricolon: the rule of 3

These devices are NOT bad. They’re the tools persuasive writing has always been made of. The problem is how frequently LLMs use them. LLMs end up reaching for them in every sentence/paragraph. Good human writers save it for special occasions

12 Aug

What is the closest to "strava for your brain" app that you actually use regularly? If you don't use it but know of one (or several), what is it?

11 Aug

Most non-fiction books should come with a companion anki cards deck

11 Aug

Replace "doom scrolling" with "boon scrolling"

6 Aug

Does anybody know why websearch tool is disabled for anthropic models on AWS bedrock? Looks like openai models got it a couple of days ago - aws.amazon.com/about-aws/wh...