AI Engineer. Data Nerd in Aotearoa. BSc Math graduate from UoA.
Auckland, New Zealand
Joined November 2021
- Tweets3.6K
- Following154
- Followers321
- Likes19K
If you can't trust a Samsung smart TV, why ever buy any "smart" TV ever again?
Instead buy the most basic dumb TV that you can, and have your own separate box you control to give it any "smart features".
为什么你花 2 万块买的旗舰智能电视，按遥控器还是卡得像十年前的安卓机？
昨天安全研究员 @log4j 把自家 3000 刀的三星电视 Root 了，扒出来的底层代码堪称年度草台班子大赏：
1. 电视后台每 500 毫秒截一次屏（连 HDMI 投屏看电脑桌面都截），打包发给近 400 家合作商。如果电视没麦克风，后台进程甚至会顺着蓝牙和回音壁去「借」一个麦克风继续听。
2. 最搞笑的来了：为什么菜单这么卡？
另一位研究员 @mike64_t 拆包发现，三星在电视那颗羸弱的 CPU 上，塞了一个叫 alex.tflite 的模型——246 万参数，2012 年古董级的 AlexNet。
这玩意每隔 2 秒抓一张画面跑推理，猜你在玩什么游戏。结果因为颜色通道写反了，它 99% 的时间什么都认不出来，纯烧 CPU。
3. 那如果你在设置里关掉「隐私追踪」呢？
代码逻辑是：先死命跑完 CPU 推理，准备上传时再看开关——发现你关了，就把结果扔进垃圾桶，2 秒后继续跑。
4. 结局更是重量级：
当号主用 OpenAI 逆向分析「自己花 3000 刀买回家的电视固件」时，OpenAI 反手一个 Cyber Abuse 把他号封了，顺带没收了剩下的 500 刀余额。
三星负责偷窥客厅，2012 年的 AI 负责烧干电视 CPU，OpenAI 负责保护三星的屎山代码不被你发现。
Sadly even with all this protection, I still became a victim of a leak:
xitter.cf/CompleteSkeptic/status…
despite the unfortunate timing, we take disclosures seriously at @typesafeai
at approximately 6:49PM this Friday evening, we became aware of a massive unauthorized leak of ALL customer noul outputs
in effort to maximize transparency, here is the link: gist.github.com/diogo149/810…
Some context for people out of the loop:
"An Indian Certificate Authority eMudhra was found to have a private key that’s literally the number 4. "
news.ycombinator.com/item?id…
David Peterson retweeted
I cannot remember any single day with SO MANY critical security breaches / disclosures, which would impact so many sites/companies/users
And have to admit that the two big labs (and especially Ant) has been warning of THIS type of thing happening, since Mythos announcement
>"Here's my hot take: your AGENTS.md and many of your skills should be deleted after every model release, or at least every couple of months, and redone. We carry so much that we think of as best practice, and we perpetuate it."
I agree. Today AI is moving faster than any new tech ever has. You can't hold on too tightly to old ways of thinking of how you use AI from last year. Or even last month!!
You must stay curious, constantly exploring new ways to use AI and pushing the limits of what you used to think AI was capable of. Otherwise you'll rapidly fall behind.
🗞️ state of agentic coding #11: the curiosity test, lisp, and labs that can't fail
ghuntley.com/toronto/
@hammer_mt even makes a case for why you should turn off entirely memories in #Claude, although I think maybe that's taking it a bit too far!!
But he still makes lots of very good points here about #ContextEngineering:
xitter.cf/hammer_mt/status/21111…
If you ever wanted to know why I obsessively start new tabs in Claude and have memory switched off, here's why. Hint: context engineering.
every.to/also-true-for-human…
David Peterson retweeted
gpt-6.1 sol examined nearly 800 mathematics papers, finding major errors in 6% and minor mistakes in another 30%. researchers estimate that 90% of its major findings are correct
i'm glad AI is now cleaning up decades of human slop
David Peterson retweeted
Accenture has told employees it will only fund 100 rows of excel per month. Meaning that employees who often use as many as 1,000 rows are scrambling to find ways to keep access.
New: Accenture has told employees it will only fund $29 per month of Claude use. Meaning that employees who often spend as much as $1,000 a month are now scrambling to find ways to keep access to the tool. A sign of the continuing crack down of AI 404media.co/accenture-tells-…
Went along to another @AnthropicAI Meetup last night (I'm a keen user after all, who makes the most of my #Claude Max 20x subscription every month!).
Hosted by @adam91holt , of thecolab.ai (photo shamelessly stolen from him, is his perspective while presenting this).
That's me on the far left in the fluro vest! 😆
(I'd just arrived on my motorbike and I was still soaking wet and freezing 🥶 💀)
Is true, if you measure the cost of AI as "per unit of intelligence" (rather than total AI spend, as for sure, companies are seeing their *total* AI costs exploding!) then the cost of AI is dropping very fast!
What will happen when AI becomes so cheap it's not even worth metering? (like how the internet is for most people today, almost nobody cares about what each megabyte of internet data costs!)
David Peterson retweeted
"Oh this? This is just an entity from the nether realm I summoned by enscribing runes on magic stones I forged in a sacred kiln, then proceeded to cast certain spells to distill the collective conscious to imbue it into the substrate. Don't worry, it doesn't have a soul."
David Peterson retweeted
"Our initial evaluation suggested that GPT-6-Astra's score was in the late eighties followed closely by Opus 5.5. But, as we started to investigate the mistakes that the models might have made, the models, especially Astra, insisted that they were right and that it was the expert-supplied answers that were erroneous. When we went back to the experts it turned out, in multiple instances, that that the experts found some subtle errors buried deep inside their calculations. As they revised their answers, Astra's score started to inch up and it finally ended up with 98% (49/50)."
Everyone is acting like this is the first time machines mogged mathematicians.
In 1976 they *lost their minds* over the first computer aided proof, quotes via new @PessimistsArc post newsletter.pessimistsarchive…
Imagine if every product manual looked like this.
Introducing Anatomy
Describe a machine in one line. Claude brings it to life.
- Interactive isometric wireframes you can rotate, inspect and explore
- WebGL shaders for fire, water, plasma and more
- Real 3D motion, with individual parts that rotate and move
Install it as an agent skill. Fully open source.
David Peterson retweeted
People need to know and understand what's happening right now
722 AI math papers from OpenAI.
53 math problems solved by DeepMind Nexus AI
(Google DeepMind’s AlphaProof Nexus solved 9 open Erdős problems, including 2 that had remained unsolved for 56 years, and proved 44 open integer sequence conjectures, with proofs formally verified in Lean).
There are many others (mostly done by individuals using the newest ChatGPT model or Claude Opus 5.5).
My point:
We're still only 1/3 of the way through this month, and we're quickly approaching the first month in world history in which AI will solve 1000+ math problems.
If the pace remains the same or accelerates (OpenAI is rumored to have 2 more big drops coming), we could reach or exceed 3000 problems solved this month alone.
This will probably be equivalent to 50-100 years worth of math progress (at normal human speed).
Reminder:
September: the number of solved problems was 10x lower (or lower).
August: ~10x lower than in September.
So we're talking here about more than 100x jump in barely 2 months.
Can you see where this is going? Soon: physics, chemistry, biology, materials science, and all other hard sciences.