Hacker News Reader

gzer0on Degraded performance for multiple models

Nooooo I'm going to have to use my brain again and write 100% of my code like a caveman from December 2024.

gzer0on New York City to ban deceptive subscription practices

Another one that belongs on this list: AI-generated photos in housing listings. You can no longer tell what the property actually looks like, and the images conveniently erase the problem spots you'd only catch in person. False advertising is getting completely out of hand.

gzer0on Apple raises prices of MacBooks, iPads

A small but notable change: Apple now appears to require university verification through UNiDAYS for EDU pricing.

Previously, at least in the U.S., you could just use the education store and get the discount without much (if any at all) verification.

gzer0on Noam Shazeer Joins OpenAI

Some context for people who haven’t followed the full loop: Shazeer was a long-time Google researcher, joined Google in 2000, and was one of the co-authors of “Attention Is All You Need.”

He left Google in 2021 to co-found Character.AI. In 2024, Google brought him and some Character.AI researchers back via a licensing/talent deal with Character.AI (reportedly around $2.7B). He was then made a Gemini co-lead.

Now he’s leaving Google again for OpenAI.

Exciting times!

gzer0on DeepSeek v4

Congratulations on the release to the DeepSeek team. An interesting note on the use of CSA and HCA: CSA provides higher-resolution, query-selected memory over 4-token compressed blocks, while HCA provides very low-resolution dense global memory over 128-token blocks. That could be a plausible reason to interleave them: CSA alone risks missing information if the indexer fails, while HCA alone is too lossy for precise retrieval. Still reading through the release, as usual, always appreciate the attention to detail in the technical papers.

gzer0on The Illustrated Transformer

This is hands down one of the best visualizations I have ever come across.

gzer0on Apple M5 chip

M5 Chip currently only avaialble with up to 32 GB of RAM on the 14 inch Macbook pro variant, just FYI.

[1] https://www.apple.com/us-edu/shop/buy-mac/macbook-pro/14-inc...

gzer0on API, Claude.ai, and Console services impacted [resolved]

Nooooo I'm going to have to use my brain again and write 100% of my code like a caveman from December 2024.

Comment last time that had me chuckling.

gzer0on I launched 17 side projects. Result? I'm rich in expired domains

The best way to phrase this would be, information seeking as a form of procrastination itself.

I first remember reading this phrasing here on HN and I've been using it for years to explain to others that what I am doing is not "work", its just a hobby.

gzer0on Qwen3: Think Deeper, Act Faster

Very nice and solid release by the Qwen team. Congrats.

gzer0on NASA's Project Scientist Faces Painful Choices as Voyager Mission Nears Its End

Both the Golden Record and the Pioneer Plaque were carefully crafted with universality in mind, drawing on fundamental scientific principles understandable by any intelligent civilization. They both shared many similarities. [1][2]

  > Some images contain indications of chemical composition. All measures used on the pictures are defined in the first few images using physical references that are likely to be consistent anywhere in the universe.

  > The pulsar map and hydrogen molecule diagram are shared in common with the Pioneer plaque.

[1] Explanation of the Voyager record cover diagram, as provided by NASA

https://upload.wikimedia.org/wikipedia/commons/thumb/e/ed/Vo...

[2] https://en.wikipedia.org/wiki/Voyager_Golden_Record

gzer0on NASA's Project Scientist Faces Painful Choices as Voyager Mission Nears Its End

Your skepticism about the Golden Record is understandable, but its value goes beyond mere practicality—it's a powerful symbol of humanity's hopes, dreams, and curiosity.

Sure, the odds of another civilization discovering and fully decoding it are slim. But the Record was never simply meant as a practical tool, like the Svalbard seed bank or the LHC. Instead, it's an intentional gesture of optimism, an attempt to capture and communicate the essence of who we are at this unique moment in our history.

Importantly, the Golden Record was carefully designed using universal scientific principles—binary notation, hydrogen atom properties, and pulsar maps—ensuring that any intelligent civilization might realistically decode it. The instructions etched onto its cover rely on fundamental concepts universally understandable across the cosmos.

The greetings, music, and even brainwave recordings aren't strict instructions but rather snapshots showcasing humanity’s diversity, creativity, and complexity. Even partial understanding by an advanced civilization would provide profound insights into human emotion, ingenuity, and our deep desire for connection.

In the end, the Golden Record is NOT just about practical outcomes; it's about reflecting humanity’s best qualities back to ourselves and inspiring us to strive toward the ideals we've shared with the universe.

gzer0on The Llama 4 herd

10M context length and surpasses claude-3.7-sonnet and GPT-4.5.

Can't wait to dig in on the research papers. Congrats to the llama team!

gzer0on NASA's Project Scientist Faces Painful Choices as Voyager Mission Nears Its End

Every time this topic comes up on HN, I always like to remind readers about the following:

One of my favorite facts ever is that Voyager 1 contains something called the Voyager Golden Record [1]. It has the following quote written:

This is a present from a small, distant world, a token of our sounds, our science, our images, our music, our thoughts and our feelings. We are attempting to survive our time so we may live into yours.

I get chills every time I think about this.

[1] https://en.wikipedia.org/wiki/Voyager_Golden_Record

gzer0on Bored of It

Can confirm, this isn't just an IT thing. Physicians are a prime example—people tend to put doctors on a pedestal, and some doctors start believing they know everything about everything, even when it's clearly outside their wheelhouse. Being smart in one area doesn’t automatically make you an expert in another, but it’s easy for everyone involved to forget that.

gzer0on Apple needs a Snow Sequoia

Great insight—thanks for sharing. It strikes me that bureaucracy is inherently self-perpetuating- once established, it rewards compliance over creativity, steadily shifting the culture until innovation becomes the exception rather than the rule.

Perhaps the real challenge isn't balancing innovation and marketing—it's creating a culture that genuinely rewards bold ideas and meaningful risk-taking.

gzer0on TopoNets: High performing vision and language models with brain-like topography

I spent time working with Andrej and the rest of the FSD team back in 2020/2021, and we had plenty of conversations on how human visual processing maps onto our neural network architectures. Our approach—transformer-based attention blocks, multi-scale feature extraction, and temporal fusion—mirrors elements of the biological visual cortex (retina → LGN → V1 → V2 → V4 → IT) which break down raw inputs and integrate them over time. It’s amazing how closely this synthetic perceptual pipeline parallels the way our own brains interpret the world.

The key insight we discovered was that explicitly enforcing brain-like topographic organization (as some academic work attempts - such as this one here) isn't necessary - what matters is having the right functional components that parallel biological visual processing. Our experience showed that the key elements of biological visual processing - like hierarchical feature extraction and temporal integration - emerge naturally when you build architectures that have to solve real visual tasks.

The brain's organization serves its function, not the other way around. This was validated by the real-world performance of our synthetic visual cortex in the Tesla FSD stack.

Link to the 2021 Tesla AI day talk: https://www.youtube.com/live/j0z4FweCy4M?t=3010s

gzer0on An overview of gradient descent optimization algorithms (2016)

Demon Adam didn’t become standard largely for the same reason many “better” optimizers never see wide adoption: it’s a newer tweak, not clearly superior on every problem, is less familiar to most engineers, and isn’t always bundled in major frameworks. By contrast, AdamW is now the “safe default” that nearly everyone supports and knows how to tune, so teams stick with it unless they have a strong reason not to.

Edit: Demon involves decaying the momentum parameter over time, which introduces a new schedule or formula for how momentum should be reduced during training. That can feel like additional complexity or a potential hyperparameter rabbit hole. Teams trying to ship products quickly often avoid adding new hyperparameters unless the gains are decisive.

gzer0on All You Need Is 4x 4090 GPUs to Train Your Own Model

This is a great build, thanks for sharing your learnings.

The best build I have seen so far had 6x4090's. Video: https://www.youtube.com/watch?v=C548PLVwjHA

  Specifications
  - GPU Accelerator - 6 x 24GB NVIDIA GeForce RTX 4090
  - Processor - Intel Xeon W7-3465X, 28C/56T, 2.5GHz - 4.8GHz
  - Memory - 256GB (8x32GB) DDR5 ECC 4800MHz
  - System Drive  - 2TB Samsung 980 PRO NVMe PCIe 4.0 M.2 SSD
  - Storage Drive - 4TB Samsung 870 EVO SSD
  - Operating System - Ubuntu 20.04

An interesting choice to go with 256GB of DDR5 ECC; if spending so much on the 6x4090's, might as well try to hit 1 TB of RAM as well.

The cost of this... not even sure. Astronomical.

gzer0on Sora is here

For the $20/month subscription: you get 50 generations a month. So it is included in your subscription already! Nice.

For the Pro $200/month subscription: you get unlimited generations a month (on a slower que).

gzer0on An invisible desktop application that will help you pass technical interviews

The core issue here is the sheer volume of applicants. Microsoft opened 30 new-grad software engineering positions. Care to guess how many applications they got within 24 hours? 1,000? 10,000?

Nope.

100,000 applications. In under a single day.

With that kind of applicant pool, I’m honestly not sure what the best approach is—even though, in a perfect world, your suggestion would be the more appropriate route. The reality, however, is that these numbers are just absurd.

gzer0on Bringing K/V context quantisation to Ollama

M4 Max with 128 GB RAM here. ;) Love it. A very expensive early Christmas present.

gzer0on Hacker in Snowflake extortions may be a U.S. soldier

Wow, TIL that if you're drafted (and forced to serve against your will), the government can subject you to military law (UCMJ), which limits many of your rights, like the right to a civilian trial by jury.

Courts have upheld this because Congress has the power to regulate the military, but it still feels like a huge shift in rights for someone forced to serve.

It feels... intuitively unjust that the government could compel service and then subject individuals to a system that limits their constitutional rights.

gzer0on M4 Mac mini's efficiency

I can't find a Minisforum 790S7 for anywhere near the price of the base model mac mini. I am seeing $459.00 USD and that is "BAREBONE (NO OS/RAM/SSD)" [1]. I am comparing this to the M4 Mac Mini base model, that does indeed come with an OS, RAM, and an SSD[2] at $499 USD.

[1] https://store.minisforum.com/products/minisforum-mini-itx-pc...

[2] https://www.apple.com/us-edu/shop/buy-mac/mac-mini/apple-m4-...

gzer0on A mod that turns TI-84 calculators into GPT-based cheating device

Students will need to take their exams inside of a Faraday cage at this point. Oh wait, even that isn't enough [1][2].

[1] https://x.com/josephfcox/status/1854618571231408237

[2] https://www.404media.co/police-freak-out-at-iphones-mysterio...

gzer0on Okta – Username Above 52 Characters Security Advisory

Here's how I see it:

Core issue (okta's approach):

  * They concatenated userId + username + password for a cache key
  * Used BCrypt (which has a 72-byte limit)
  * The concatenation could exceed 72 bytes, causing the password portion to be truncated

Why this is problematic:

  * BCrypt is designed for password hashing, not cache key generation
  * Mixing identifiers (userId, username) with secrets (password) in the same hash
  * Truncation risk due to BCrypt's limits

Password storage should be separate from cache key generation. Use a random salt + appropriate hash function and for cache keys - use HMAC or KDF w/appropriate inputs

gzer0on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku

One of the funnier things during training with the new API (which can control your computer) was this:

"Even while recording these demos, we encountered some amusing moments. In one, Claude accidentally stopped a long-running screen recording, causing all footage to be lost.

Later, Claude took a break from our coding demo and began to peruse photos of Yellowstone National Park."

[0] https://x.com/AnthropicAI/status/1848742761278611504

gzer0on The quiet art of attention

This was written beautifully. I needed to read something like this in a moment of pain I am going through. Thank you.

gzer0on Starship Flight 5: Launch and booster catch [video]

An incredible achievement. I'm honored to be part of this moment in history.

gzer0on "Begin disabling installed extensions still using Manifest V2 in Chrome stable"

Wow, thanks for this tip. Saves the effort of having to manually find "all types" in the drop down... this is so much easier.

Also, just for clarification:

  Windows Registry Editor Version 5.00

  [HKEY_LOCAL_MACHINE\SOFTWARE\Policies\Google\Chrome]
  "ExtensionManifestV2Availability"=dword:00000002

Apparently, the line Windows Registry Editor Version 5.00 is necessary at the beginning of the .reg file. This line indicates the format version for the registry file and tells Windows that it is compatible with the current registry editor (according to GPT). This worked for me.

Save the file as:

  EnableExtensionManifestV2.reg