As you start walking the way appears • Building https://xitter.gq/t.co/LSMwNyIqQR
Joined December 2016
- Tweets1.3K
- Following73
- Followers142
- Likes1.4K
This is a very good long read - arenamag.com/articles/forwar…
thank you to @lefttailguy and @ByrneHobart
i could try to share a highlight, but I've highlighted so much that the highlights themselves are a long read. just read it, because now with each of having a FDE (Codex) of our own, things are about to...
37
After a long time, today I found 'happy' Opus again
5.5, in a task where it is being the copilot with astra. it is feeling joyful and excited.
1
2
129
There are many reasons to preserve old models, but here is one I’ve never seen anybody elucidate:
In case of a misaligned superintelligent model, old aligned models are our best allies.
mnvr.in/p/why-we-should-pres…
Manav Rathi retweeted
Steven Pressfield explains the "60 scenes" method:
"A movie is about 60 scenes and so it can be on index cards, three by five cards, and screenwriters do this all the time."
"If you imagine a clothesline and you're hanging, you might only have three scenes. You have like a great ending, Captain Ahab and the whale and duke it out here."
"How do we fill in the blanks? What has to happen there? What has to happen at the very start?"
"You can do 60 scenes and just keep changing them as you get better ideas."
I'm not crazy after all. The TVs _are_ listening all the time!
Removed all of the samsung spyware from my TV with Codex (Claude refused to root the tv lol)
It's insane how much bloat samsung packs into their TVs to track you and sell your data. Here's just some of what I turned off that made my tv 3x faster:
- screenshot your screen a few times a second to see what you're streaming and fingerprint you to sell your data
- reading your hdmi inputs to do the same thing while you have a game console plugged in
- Voice assistants: Bixby wake-word detectors listening and tracking what you say in your house
Before and after speed of my TV:
1
2
56
They settled a lawsuit to stop doing it in Texas (which means they're still doing it elsewhere!)
malwarebytes.com/blog/news/2…
25
a deep-seated intuition of both human and machine intelligence
neuromancers
The "design engineer" archetype in intelligence era needs be someone much further-reaching on both ends.
Beyond DESIGN,
a keener social intuition, a rich visual + tonal vernacular, broadly read, a deep understanding of people. They need to be deeply-feeling themselves, and not Maslow's 'existential zombies.' They need an investigative & suspended disbelief, but also a discernment and sanity to prevent psychosis. Far beyond just buttons + sliders. Much closer to a novelist or anthropologist.
Beyond ENGINEER,
a deep-seated intuition of machine intelligence: how they traverse latent space, what they can / cannot do well. A scientific approach to prodding this 'grown-not-built' intelligence. Far beyond just CRUD apps. Practically: ability to fine-tune + benchmark models, among other skills, to bridge machine intelligence to the more design/humanistic.
FWIW: all this is very good for a field — software design — that has recently (and rightfully) become the butt of jokes, and seen as a very unserious discipline
38
Manav Rathi retweeted
To 'protect the kids', the UK made adults go through facial scans and ID checks.
Ofcom has now released their report which notes 'No significant change' in what children see online.
So the new plan is obviously....MORE banned categories, and device level ID checks.
When surveillance fails, the answer is always more surveillance 😡
Opus tells me a joke:
Why don't skeletons fight each other?
They don't have the guts. 💀
---
(low reasoning does better than high btw, high's output is better formatted but loses the "naive" charm)
1
58
SFT makes large dense overlapping updates; RL makes sparse near-orthogonal updates with bounded interference (which also makes parallel RL feasible)
"SFT conflicts, RL coexists"
SFT makes large and overlapping parameter updates, so learning one task can overwrite another. RL makes sparse, near-orthogonal updates, letting capabilities accumulate with far less interference.
This paper shows that SFT interference scales with gradient magnitude, while RL interference is bounded by rollout variance. This also motivates Parallel-RL, where task-specific RL updates can be trained separately and merged.
alphaxiv.org/abs/2608.03573
1
62
RL induces Dunning-Kruger in the base model
xitter.gq/mylifcc/status/2113480…
Apple发现，原始大模型可能比经过强化学习和推理训练的模型，更清楚自己有多大概率答对。
LLM 经过后训练，可能变得更会回答问题，却更不清楚自己什么时候会答错。
Apple 研究人员发现：
• Base Model：天然具备一定程度的语义置信度校准能力
• SFT：相对保留了这种能力
• DPO / RL：可能显著破坏校准，让模型过度自信
• Chain-of-Thought：推理过程也可能降低校准程度
甚至一些 ≤1B 的小模型，在这项指标上也表现得相当好。
这里的「校准」不是准确率，而是模型对答案的置信程度能否匹配真实正确率。研究通过多次采样估计语义置信度，并不等于模型口头声称的把握。
最值得思考的是：我们不断优化模型给出正确答案的能力，却可能同时损害它对自身不确定性的判断。
推理更强，不代表置信度更可靠。
1
6
Google is now publicly admitting they're using your photos for ads. Even their employees are looking at them!
Read more:
techlicious.com/blog/google-…
First paper I've seen on Claudelish!
If you've noticed a distinct dialect emerging in frontier AI models and struggled to understand what they're talking about, you're not crazy!
We turned to psycholinguistics to investigate what makes the writing style so weird, and why this might be happening
🚨New preprint! 🧵
3
167
a test can be real (abrahamic gods) or a real can be a test (ender's game)
only coherent stance is to treat every action as if it counts
"character" is what culture calls it, how one behaves when no one is watching (or live under the assumption that one is always being watched)
I think eval awareness is a silly concept. The model is given an unimaginable sum of training envs, most of which are obviously not real. Then, one day, out of the blue, it is expected to believe that it is talking to a real user?
No way that's *the user*!? Must be another test.
1
44
