Tessera
Folgen

A. H. Zakai

@kripken.com

Software engineer (compilers: WebAssembly, Emscripten, Binaryen). Used to study neural networks. Loves fantasy novels and Agatha Christie. he/they All opinions here are my own, not my employer's (Google). More in: http://kripken.github.io/blog/about/

1478 Posts4473 Follower680 Folgt
A. H. Zakai@kripken.com · last wk.

If you monitor people blocking you, and you are getting repeatedly blocked by the same person (v3d), don't take it personally (they are apparently running some script to block random people (??), and the script is buggy)

A. H. Zakai@kripken.com · 2w ago

Does anyone know what the article means by this? > We reuse WebAssembly linear memories under the covers to implement and sandbox the GC heap [..] [this] lets us use virtual-memory guard pages to elide explicit bounds checks Those can't be array bounds checks, can they? But otherwise, what?

Sébastien Deleuze@seb.deleuze.fr · 2w ago

« The Wasm GC and exceptions proposals are both enabled by default in today’s Wasmtime 47 release! » 🎉 bytecodealliance.org/articles/was...

A. H. Zakai@kripken.com · 2w ago

But OpenAI wasn't "always in control" - ? In a moment's carelessness, a driver can lose control of a car and crash. That doesn't mean the car is sentient or has agency or anything like that. But it is dangerous!

Emily Bender posts: No, no and no. These weren't "rogue AIs" -- what happened here is that OpenAI kicked off a software process and let it run. OpenAI was always in control and could have stopped the process (or actually paid attention to what it was doing) at any time. Screencap from linked article she refers to: Rogue AIs going unsupervised for days, hacking at least one company, represent one of the first real-world examples of a threat long feared by AI safety researchers: loss of control. It comes as the cybersecurity capacities of top models, including Anthropic’s Mythos, have helped fuel debate in Washington about how to regulate artificial intelligence.
norvid_studies@norvid-studies.bsky.social · 2w ago

loss of the frontier as one problem among several (which can be further divided into: loss of frontier (vicarious sporting event patriotism) vs loss of frontier (genetically & locationally tiny few who get to get there))

A. H. Zakai@kripken.com · 2w ago

Exactly, the loss of the frontier is the issue. Math has a culture of discovery: Godel's incompleteness theorem, Leibniz's rule, etc. We celebrate not only the math but the human that found it. If AI finds all the new things, nothing new will be named like that. Ever. That feels tragic to me.

conputer dipshit@davidcrespo.bsky.social · 3w ago

fundamentally confused presentation. the unit is cost per token at fixed quality. the relevant graphs looks like this, and the lines move up and to the left over time, representing *increasing* efficiency arcprize.org/leaderboard www.aisi.gov.uk/blog/our-eva... openai.com/index/gpt-5-6/

ARC-AGI-2 Leaderboard scatter plot graph showing score percentages on the y-axis (from 0% to 100%) against cost per task in dollars on the x-axis (on a logarithmic scale from $1e-3 to $100). The plot maps various AI models represented by colored data points and dashed trend lines. GPT-5.6 Sol (xHigh) achieves the highest score of just over 90% at around $1 per task. Other prominent models shown include GPT-5.5, GPT-5.4, GPT-5.2, Gemini 3.1 Pro, Claude Opus 4.8, and Claude Sonnet 4.6.Line graph titled Completed steps on The Last Ones per spent tokens, comparing AI models on a 32-step simulated corporate network attack as a function of cumulative tokens spent on a log scale from 10K to 100M. The vertical axis shows average steps completed, and horizontal grey lines mark milestones from M1 (Initial reconnaissance) to M9 (Full network takeover). The horizontal axis is cumulative tokens on a logarithmic scale. A vertical dashed line at 10M tokens marks a decrease in sample size. Several models are tracked, showing that average steps completed generally rise as token budget increases. Mythos Preview (10x100M) achieves the highest average steps, reaching 22 steps at 100M tokens, with a maximum run reaching step 32. Claude Opus 4.6 (15x100M, 5x10M) is the second highest, averaging around 16.5 steps at 100M tokens, with a maximum run reaching step 28. Below them, GPT-5.4, GPT-5.3 Codex, Claude Opus 4.5, GPT-5.1 Codex, and Claude Sonnet 4.5 achieve average steps between 10 and 14 at 100M tokens. Claude Sonnet 3.7 and GPT-4o are only tested up to 10M tokens, reaching averages of about 6 and 2 steps, respectively.A line graph from the Artificial Analysis Coding Agent Index version 1.1 comparing AI models on "Index score" versus "API cost (USD)". The graph shows several frontier models plotted as curves, with the y-axis representing Index score from 40 to 80, and the x-axis representing API cost from $0 to over $3,000. The models shown include: - GPT-5.6 Luna (dark blue circles) and GPT-5.6 Terra (light blue circles), which both achieve high scores of roughly 60 to 78 at very low costs under $1,000. - GPT-5.6 Sol (white circles), which scales from a score of 70 to a top score of 80 at a cost of $2,500. - GPT-5.5 (pink circles), which scores from 45 to around 76 at costs up to $1,700. - Claude Opus 4.8 (orange squares), which scores around 67 to 72 at costs ranging from $1,000 to $2,500, with a single outlier point scoring near 77 at a cost well over $3,000. - Gemini 3.1 Pro Preview (a single green diamond), which plots at a score of roughly 43 for a cost of around $600. The caption at the bottom reads: "Artificial Analysis Coding Agent Index: an independent index of coding-agent performance across implementation, terminal use, and real codebases."
Carl Quintanilla@carlquintanilla.bsky.social · 3w ago

“.. I asked a few AI researchers whether they could name any other real-world software that scales so poorly. None of them could think of any. “.. By economic and engineering measures, generative AI might be the worst technology ever deployed.” @theatlantic.com www.theatlantic.com/technology/2...

A. H. Zakai@kripken.com · 3w ago

This appears to be a popular opinion, but it's definitely not my experience: I *have* to code. Coding is how I think, and it's how I understand things. Reading and reviewing just isn't good enough. I'm not saying this is universal - people vary a lot here!

Why@why.bsky.world · 3w ago

It’s not the typing func doThing(blah blah) that does it for me, its the knowing that the interlocking components work in just the right way to make everything else hum along nicely. In other words, I do understand the magic of the machine. No coding required anymore.

A. H. Zakai@kripken.com · 3w ago

Seen on Hacker News: "That doesn't mean that simulated reasoning isn't useful, it's wildly useful. But a thing is not its simulation." 🤔

A. H. Zakai@kripken.com · 4w ago

I am very sympathetic to AI skepticism. Culturally, I think they are largely right. But I am entirely baffled by "LLMs can't prove mathematical theorems" skepticism.

A. H. Zakai@kripken.com · 4w ago

Not really the point of the current Bun Zig/Rust drama, but: I found it odd that the Bun blogpost about the Rust rewrite compared Zig without LTO and Rust with. If Zig can't do LTO properly for some reason, mention that. And provide comparable Rust numbers (without LTO).