Max Van Kleek–Assoc Prof @CompSciOxford, a16z speedrun sr006, Scientist/Builder/Founder, Human-AI Interaction, Future of Work. 🇺🇸/🇬🇧 🌈

Oxford/London + SF
Joined May 2007
electronic Max retweeted
ONE WEEK LEFT The final day to apply for our next residency is Sunday, Nov 1st. If you have the makings of a generational founder, don't hesitate. It's time to build.
10
6
4
39
3,420
electronic Max retweeted
Kiddo is already showing a strong understanding of dependencies (xkcd 2347)
24
240
8
4,464
49,604
electronic Max retweeted
Now kiss
137
2,196
164
30,935
645,803
i love that influencers are just doing extreme user testing for @Apple for free.
1
150
This is the first time I've had an AI song pass the "make you dance in your kitchen" test. Didn't even realize this was an AI song until after. Absolute banger song and dance.
This will be the soundtrack of my sleep.
30
28
3
776
178,081
british summer time being over means there is only darkness for the next 6 months
1
6
309
if it didn't pass the british winter sun test it wouldn't survive anything - those 6 feeble photons that somehow made it through the thick grey layer of gloom
iPhone Duo passes the British sun test
4
270
electronic Max retweeted
Replying to @tamika47488
6
5
1
342
96,939
electronic Max retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
134
781
143
6,395
762,854
electronic Max retweeted
design wise i have never been more uninspired by anything apple has released in the last 20 years. it’s not just the 'apps aren’t adapted yet' problem (and the irony is that in most cases the apps that haven’t been adapted look better than the ones that have) or the usual first iteration issues. it’s a series of fundamentally weird design choices that make it impossible to even come up with nice layouts. there are screens where nothing aligns with anything, buttons floating around randomly. forget about symmetry. like, forever. it’s not even a concept anymore. not a single screen on the main display has made me think, 'oh, that looks neat' (or balanced, or at least makes sense from an IA perspective). everything is just...random. but yeah, a nice opening animation (that you barely ever actually see), a cool combo status bar (that most of the time just sits there all alone, taking up 1/5 of the screen’s width), and a nice click when opening or closing it (no comments here, the click is actually nice).
61
56
18
887
111,451
I absolutely do not want a deep uncanny valley voice AI. Completely unnecessary for me. I'm completely fine with talking to something a little bit Robotty.
1
2
149
this is well known in computer security already; models exploited well understood side channels such as from spectre and meltdown
“there are unfixable bugs in hardware that are exploitable by AI models directly speaking machine code”
3
285
the use of hardware side channels is definitely impressive
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. openai.com/index/securing-ou…
146
"It will come alongside a £60m investment in what it coined a "National Cyber Shield" which it said would use agentic AI to boost cyber-defences, and £70m funding towards a National Exposure and Real-Time Vulnerability (NERV) Centre." bbc.co.uk/news/articles/cqdj… Great! How about for protection national infrastructure, including utilities, transport, comms?
1
278
electronic Max retweeted
นักพัฒนามันทำแต่แอปแบบนี้55555 ชอบ
49
383
88
5,243
2,497,838
electronic Max retweeted
i thought vibe coding was just when you lit a couple of candles, some incense, some Bach in the background, big slippers and a dressing gown on... just cosy vibes coding
6
1
38
865
some beautiful interaction design here. my question, where do you put the apple pencil when you're not using it?
Taking notes on the iPhone Duo is so satisfying… especially when it looks and feels like a real notebook that truly belongs to you. With TypeDraw you choose what happens when you “open” your book, and you can personalise it with custom colours, fonts and even your own handwriting. For everyone lucky enough to have their hands on a Duo today: TypeDraw is available on the App Store now. For everyone else: it also looks and works great on regular iPhones 😄
5
9
647
electronic Max retweeted
iPhone Duo is here! Excited to have worked on the design of our open/close transition. It’s designed to make the two displays feel like one seamless experience by perceptually bending the focal plane, not the content, so that it feels like opening a window onto a singular view.
266
498
65
8,230
493,042
Never trust anyone who uses Outlook Reactions.
1
7
318
technology would make us smarter they said
THE FUTURE IS HERE
135