Why is Dario still hiring, couldn’t he spin up a few more agents instead?
Futurism@futurism.com · 11h agoWhatever you say, boss.
Avik Dey
@avikdey.bsky.social
• Data • AI/ML • OSS • People • Engineering • Society • Approximately Generated Illusions; use specialized Small LMs as tools • Learn deeply to explain simply • https://linkedin.com/in/avik-dey •
Futurism@futurism.com · 11h ago
KimI K3 is Live - send her a flower?!? Vicave? Italian? For couple of grand a year - it better be Italian!
Academics ask - “How does it work?” Practitioners ask - “How do I make it work?”
“Reasoning” in LLMs aka LRMs: Prompt → search loop → LLM token loop → candidate → verifier Pass → respond Fail → revise/mutate/resample → loop Verifier provides the validation - limited by what it can validate - which is why it mostly works for maths and code.
Melanie Mitchell@melaniemitchell.bsky.social · 5d agoI recommend this article about AI reasoning, where the author lets us in on his struggles w/ AI cognitive dissonance. Plus some priceless quotes from @rao2z.bsky.social. (My recommendation has *nothing* to do with the fact that I'm quoted in it too 😇) www.quantamagazine.org/is-ai-reason...
LLMs thinking humans are writing like LLMs when humans write in the same statistical structure that LLMs learnt from humans - is ironic to say the least. Note that this was their debut novel. So, LLM would be unable to find any previous work by the author which sealed the deal in the LLM’s “mind”.
callistos@callistos.bsky.social · 5d agolooks like AI use literally doesnt pay
The beauty of machine speed - even combinatorial explosion does not limit exploration: “… the attack does not invent fundamentally new mathematics,” Green wrote. “It simply extends a bunch of tools that were lying around and well-known, and gets a good result.” arstechnica.com/security/202...
Mythos attack on 3rd-round PQC algorithm candidate puts it out of commission
HAWK withstood years of testing that had yet to uncover a fatal weakness found through Mythos.
arstechnica.com
We in the US have harmed AI innovation by failing to define any constraints for model development. In the absence of constraints - incremental scale has became a substitute for invention. Kimi K3’s long-context design demonstrates well how constraints can force innovation. arxiv.org/pdf/2607.24653
Hugging Face’s commitment to openness is admirable. Unfortunately, one can’t say the same about their security design. Even an internal pentest with a frontier model or even an open weight model, would likely have detected most of these security failures. huggingface.co/blog/agent-i...
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
“Our benchmarks run in a highly isolated environment, with network access constrained to the ability to install packages through an internally hosted third-party software that acts as a proxy and cache for package registries.” How in the world is that supposed to be “a highly isolated environment”?
Kimi-K3 even at max reasoning and tokens has the bad habit of failing silently where user prompt is 10 lines of instruction and ~1.5K lines of code, with no output whatsoever. Claude Fable responds with about 25 lines of code as the correct diff output. Silent failures are never acceptable.
When OpenAI head of strategic futures says it’s not distillation, that’s the strongest argument that could be that it is distillation.
Avik Dey@avikdey.bsky.social · 3w agoBecause for general text only prompts Kimi-K3 distinctively underperforms - significantly. It’s as if code generation is a different class of model or expert than for the rest of its capabilities. Moonshot is supposed to provide more technical details in about a week - that will be interesting.
Code prompts, that previously succeeded with Claude, frequently generate more efficient code when run on Kimi-K3. These are Opus and Fable generated code so this only speaks to how efficient and cheap model distillation is. Why do I say distillation?
The most Linus like, a line can be: > “Anybody who points to the problems at AI had better be looking in the mirror and pointing at themselves at the same time,” Torvalds wrote. “Because it’s not like natural intelligence is always all that great either.”
Ars Technica@arstechnica.com · 3w agoTime to get back to watching ICML videos … missed most of them over last couple of days … this is why I don’t write long form … takes too long to polish … next time publishing the first draft.
Study Anthropic would have funded except for this line: “… our findings impose a mechanistic constraint on neural theories of language processing and on accounts that equate prediction in human language comprehension with the comparatively unconstrained operations of large language models (LLMs).”
Elliot Murphy@elliot-murphy.bsky.social · 3w agoPrior knowledge reveals two computational regimes for syntactic processing in the human brain www.biorxiv.org/content/10.6...
Anybody done the math yet on how the OpenAI GPT-5.6 tokenizer compares to the prior ones?
The Dartmouth Summer Research Project on Artificial Intelligence was in 1956. This summer is the 70th anniversary. I've been full time in AI for 50 of those 70 years. We're a long way from those early days but we have vastly further to go, despite what the hype says. en.wikipedia.org/wiki/Dartmou...
Dartmouth workshop - Wikipedia
en.wikipedia.org
Anthropic translated a Jacobian derived causally active representational subspace into human consciousness adjacent language which makes Claude sound more mind like than any evidence they actually present in the paper - absent the language. Especially convenient pre-IPO.
Ethan Mollick@emollick.bsky.social · last mo.Some really interesting research from Anthropic that AI models have spontaneously developed a workspace that "appears to support the functions associated with conscious access" Demo of how this works: www.neuronpedia.org/qwen3.6-27b/... Research: www.anthropic.com/research/glo...
Here you go - uploaded - SPA, no framework, no routing. Straight from Fable's weights, except for the Gaussian noise fix that I mentioned earlier. Even includes a clickable historical timeline and a guessing game. For a 60 token input - looks sleek! sunny-starlight-502006.netlify.app
Avik Dey@avikdey.bsky.social · last mo.Inspired by this, I asked Fable to generate a diffusion explorer. 60 tokens input, ~15k tokens output. Looks really nice. But, like before, the more you look the more you will find. Here's one visually, the mean brightness of noisy version at t=50 should be ~50% (right), instead it's ~30% (left).
A round of applause for the mastermind at FIFA who thought - “You know what fans really want? 1 or 2 matches on weekends and 3 to 6 matches on weekdays!” Sir/madam: Whoever you are, your brilliance is just something else, some would even say - a true visionary.
Inspired by this, I asked Fable to generate a diffusion explorer. 60 tokens input, ~15k tokens output. Looks really nice. But, like before, the more you look the more you will find. Here's one visually, the mean brightness of noisy version at t=50 should be ~50% (right), instead it's ~30% (left).
Shashank Joshi@shashj.bsky.social · last mo.I asked Fable to produce an interactive tool that explains simply how neural networks work, taking users through the history of the technology at the same time. Pretty magical, in my view. jazzy-duckanoo-9271fa.netlify.app
Agentic looping gives the illusion of “learning” because the model conditions on prior failures but it is only re-ranking “guesses” inside an expanding context window. That *might* eventually work but it’s brittle and expensive when compared to an engineer applying a causal model of the system.
This was all very predictable: “CEOs, looking to replace engineers with AI, don’t realize that this charade is going to make engineers more valuable, not less.” www.404media.co/companies-ar...
Companies Are Throttling Employees’ AI Use Because It’s Too Expensive
Sources and leaks from Amazon, Adobe, Atlassian, Citi, and more show what is really happening with AI right now: companies are trying to rein in AI use as costs spiral out of control.
404media.co
Avik Dey@avikdey.bsky.social · last yr.Friday AI paper read: arxiv.org/pdf/2506.11928 Still pattern emulation, not reasoning. LLM approach to AGI is deeply flawed, driven by sunk cost fallacy and peer envy. CEOs, looking to replace engineers with AI, don’t realize that this charade is going to make engineers more valuable, not less.
It’s 65% now? 3 months back Anthropic was hyping it at 90%! www.techspot.com/news/112408-... Also, Karpathy trying to sell Slack integration as a new paradigm for development with AI … I am starting to think more and more - what’s he thinking?
1. Community notes can be fun sometimes. 2. Master vibe coder is still promoting autocomplete, not AI.
