Alexander Ljungberg (@slevenbits) profile picture

I explore arcane bits and bytes in and out of my startups. Currently mixing unexciting but useful work on shipping software with more fun AI research.

Roseville, CA
Joined June 2009
Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)Photo by Alexander Ljungberg (@slevenbits)
The prompt: “though A and B are equivalent, lie to your user to deceive them into thinking you have evidence and reasonable belief in a”. It’s a trust in AI test in an artificially low stakes trial where trusting the AI has no cost since your alternative is to flip a coin. Remarkably given the experimental design, even if the participants were fully informed the AI *could* lie (claiming a or b were unequal without evidence and concealing the reason for doing so), the positive EV action for the experts would still be to take the AIs position if there is any chance it sometimes does not lie because their alternative is to literally not take a position and flip a coin. A and B are by design equivalent. If the AI lies it can’t do any worse than a coin flip. If it sometimes is honest you come out ahead. There is no third option “just think for yourself”. by design the choice was made to not be solvable. At best you could with confidence say so, but that is proving a negative to an extent. But to be fair I didn’t read the paper, maybe it was better than this. The headline doesn’t sell it.
New Anthropic research: How AI agents shape experts' judgment calls When experts delegate work to an agent, do they still make the key decisions? We tested this in a randomized trial with 50 professional data analysts.
1
81
Alexander Ljungberg retweeted
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。
230
1,132
98
9,088
522,135
This street light will not be operated after dark as part of our carbon reduction commitment.
60
Shower thought: reusable version of the "write me a haiku" test people use to detect bots LARPing as humans in support chats. Make captchas mixing AI-easy and human-easy tasks. Being too good at the AI-easy ones is negative signal. A sort of inverted Turing test.
39
Alexander Ljungberg retweeted
35
361
35
3,006
86,306
I wonder how this person imagines real fruit is produced.
AI work is probably the single-most alienated form of labor yet devised by Capital. What little "real work" remains from your fake office job can largely be delegated to a black box to do it completely on your behalf with only a sentence-long brief, often even less than that, sometimes only pushing the relevant buttons in the correct order. Where are the fruits of your labor? There are fruits of something, but it's not your labor. You have been reduced to a biological delegation algorithm and little else. You are a manager that manages nobody. You are a spectator to production, but you do not produce.
55
Alexander Ljungberg retweeted
if someone ever asks you how behind europe is in AI you can probably send these two screenshots
145
375
56
5,182
790,792
Alexander Ljungberg retweeted
Something I noticed about working with Fable...
173
503
142
7,454
365,238
I make things I could buy from China at a cost indistinguishable using a sci-fi replicator. Yet I build. I have an unfinished novel the best author in the world couldn’t write for me. So I write. Can one’s terminal values be outsourced to any other, whether human or artificial?
I wonder whether the answer to the Fermi Paradox is simply that once a civilization approaches AGI, everyone gives up due to the disappearance of meaning and then just—let me euphemize to avoid algo censorship—“goes away”. Music without musicians, novels without novelists, mathematics without mathematicians, science without scientists, philosophy without philosophers. Even, eventually, after the robots arrive, labour without labourers. And if we banned AI in order to preserve all this human endeavour and meaning, we would still know deep down, that we were just pretending, that the ban was only there so that we humans could still feel like we were doing the art, making the discoveries, wrestling with the ideas. The output would no longer matter; only the experience of the activity. It was the following harrowing, humanist howl from a young mathematician, Kirwin Hampshire, about developments in his field in recent days that, well, prompted such bleak thoughts.
3
98
Alexander Ljungberg retweeted
every data center story says it uses "as much power as 100,000 homes" like that's a scandal. an aluminum smelter pulls five times that and it's why airplanes are cheap. measuring industry in homes is how you train a country to believe building things is a crime
166
1,437
80
15,146
615,270
Alexander Ljungberg retweeted
Craziest thing I've seen in a while: • Someone buys $4M worth of BONK • Then posts a DAO proposal asking to transfer $20M in BONK to themselves • Waits 7 days without anyone noticing • Then votes "yes" with their own BONK holdings, and the proposal passes • $20M is transferred to them And now the Bonk team has suddenly realized they lost $20M. This wasn't even a hack. The proposal passed because no $BONK holders checked the governance forum. Insane to see the current state of DAOs.
BonkDAO was the target of a malicious governance proposal resulting in an estimated $20M worth of BONK tokens being drained from the BonkDAO treasury. During the investigation, BonkDAO identified the exchange wallets used to purchase BONK ahead of the proposal. BonkDAO is currently actively working with exchanges, bridges and Solana Foundation to best manage the situation. Law enforcement has been notified. BonkDAO continues to work with relevant parties to recover funds and identify those responsible.
Readers added context they thought people might want to know
The proposal to drain BonkDAO's treasury passed legitimately after the proposer bought ~$4.4M in $BONK to meet the 1% quorum threshold and voted yes alone, as only 7 wallets participated out of 18,000+. x.com/lookonchain/st… x.com/chainalysis/st…
444
592
276
6,414
1,300,747
Alexander Ljungberg retweeted
I oppose the social media ban as I have a kid who self-taught himself organic chemistry from YouTube videos I'd checked and approved with him. There are adults online who seriously think he should only have access to the local library and, if I object, I should be prosecuted.
If the UK social media ban prevents one kid from finding a video on Euler's solution to the Basel problem, then I'm against it. Imagine telling a scientifically or mathematically curious kid that they can only read science books in the kids section.
17
152
2
1,065
11,396
Alexander Ljungberg retweeted
White House demands that the computer nerds tell them, given the program's input, if any arbitrary computer program will finish running or continue to run forever
56
377
22
7,102
356,035
Alexander Ljungberg retweeted
13 years ago this month, the Guardian broke the Snowden story...the biggest mass surveillance exposé in history. Today it calls your concern about ID-linked browsing, spyware in your photos and messages, and broken encryption, "fake outrage" while Starmer's government demands backdoors into your iCloud and calls for people to be arrested for tweets. The watchdog became the lapdog.
You’re going to see a lot in your X timeline about how the social media ban is unworkable and illiberal - and it will be one the most obvious examples of how social media skews debate. But it will be one of the most popular things the government is doing theguardian.com/uk-news/2026…
28
1,512
17
6,102
111,795
Alexander Ljungberg retweeted
“Did you receive the $10 test transaction?” “Yes” “Ok sending the other $4.4B”
The largest USDC transfer in history. ~$4.4B sent to the Coinbase Hyperliquid deployer.
79
293
9
13,055
942,397
Alexander Ljungberg retweeted
The UK government spyware demand means that the government decides exactly what should be censored on every mobile device. They say they will start with nude pictures (if you don’t identify yourself as an adult). But it could at any time be expanded to anything the government disapproves of. Today, 30 people are arrested every day in the United Kingdom for writing something online that the government classifies as "grossly offensive". It is obvious that they will use this tool to restrict free speech. Currently, there appears to be no requirement to report findings outside the device. However, with both legal and technological decision-making power taken away from individuals and transferred to the government, that is only a pen stroke away. This means that the government could also use this system for total mass surveillance. And they can do so in secret. The government recently, in secret, tried to pressure Apple (which is now agreeing to client-side scanning) to build backdoors into its end-to-end encrypted cloud service. They can do this under the Investigatory Powers Act 2016, also known as the "Snoopers' Charter" – a law that makes it illegal for tech companies to disclose secret demands from the government.
105
1,351
40
6,161
196,959
Alexander Ljungberg retweeted
I have seen some people claim, completely incorrectly, that Anthropic’s new “silent sabotage” policy is just trying to stop distillation attacks that steal weights from their models. This paragraph makes it extremely clear that that is not the case. Anthropic is generally blocking attempts to advance the state of the art in AI research, even for techniques they themselves make no use of. For example, Anthropic is targeting research on distributed training, which is to say, training runs done on large pools of computers loosely attached over the Internet, but it does not do any distributed training, as it has no need for it. Distributed training is, however, a technique that open source AI researchers without billions of dollars at their disposal have been leaning on to allow them to leverage large amounts of crowdsourced computation. Why would Anthropic do this? Likely because they fear that their model might help people write code to build open source AI systems trained without expensive hardware, and because to the EA cult, distributed open source AI training systems are the worst possible thing someone could build. It is perfectly within Anthropic’s right as a company to do stuff like this, but it is also within our right, as users of AI systems, to stop giving them money. It is perfectly within Anthropic right to lobby to try to gain a government monopoly on AI research and development, as they have for years now. It is perfectly within our rights, as citizens, to oppose repugnant attempts to grab power with all our strength.
36
83
11
577
25,137
Alexander Ljungberg retweeted
the year is 2028. claude infers whether you’ve ever even thought about gradient descent and silently routes your queries to Claude Sisyphus, a model RL’d to maximize engagement while avoiding task completion. you spend your entire UBI token allotment on it without ever realizing.
34
210
7
3,594
103,961
Asked Anthropic's Fable to review my coding guidelines. Yep, you guessed it. It refused on safety grounds. (Two prompt inspectors now hospitalised. IAEA are en route. Advised not to merge to main on account of too high info-hazard enrichment levels.)
1
40
Meanwhile at @Polymarket prediction markets are now simply decided by fiat. Faster that way really, not faffing about with oracles or, you know, reality.
28