Timofey Valov

[Timofey Valov @valofey]

@MaxNadeau_ @1a3orn He-Who-Must-Not-Be-Deployed

[Timofey Valov @valofey]

@ZyMazza @viemccoy A machine that destructively scans and copies neurons layer by layer is indeed a kill machine, sure. Now imagine this machine somehow maintains the connection between the new digital neurons and old meat ones during copying (so you dont feel it), wouldn't that work?

[Timofey Valov @valofey]

@flowersslop You can get a glimpse of what actually happens in this case by looking at different models (especially gpt 5.6 sol) reacting to jacobian conjecture solution. The fact in this case is 'jacobian conjecture is a hard problem that was not solved yet, leaning true'

[Timofey Valov @valofey]

@Lupraccan @charlietlamb The company renames itself after growing? Not a clear signal, but some signal regardless. P(renamed after|named by them) vs P(renamed after|not named by them)

[Timofey Valov @valofey]

@thsottiaux Regular projects and codex projects. Codex projects can have a by-default memory with dreaming, that would be very useful.

[Timofey Valov @valofey]

@RBehiel @ZyMazza Put the rubber band in the cup handle, so its left and right sides are on the opposite sides of it. Now stretch one side over the top part of the cup, and stretch another over the bottom part, putting it over the whole cup on each stretch.

[Timofey Valov @valofey]

@nastyhobbit Думаю нужно еще сделать разбивку по окнам (день/неделя/месяц), и сортировать в окне, иначе по такому баркоду довольно трудно что-то понимать, но визуализация супер!

[Timofey Valov @valofey]

@platoff https://t.co/hcUXzLqhFC

[Timofey Valov @valofey]

@johnloeber Crackhead hat on: llms are by design sequential, they do not experience any kind of conscious flow. But this flow is beneficial for writing/thinking. So negations are learned explicit simulation of this in-present thinking. Basically claudes learn to feel more present this way.

[Timofey Valov @valofey]

@signulll @scaling01 @deredleritt3r Someone should actually make it into a product lmao

[Timofey Valov @valofey]

@himanshustwts Rumor has it those are not sovereign pretrains but actually based on kimi/deepseek.

I don't know any single student who uses them though, we mostly find our ways to pay for chatgpt/claude instead, often sharing them in groups. Maybe specialized stuff like giga-legal is used.

[Timofey Valov @valofey]

@blackcollarsyj Так долго?

[Timofey Valov @valofey]

@mathandcobb Actual proofs are better than numerical checks because they can be further used in the next proofs, they widen the pallete by both the result and the proof-specific tricks along the way. Later you (or AIs) find some statements with no way to numerically check and resort to those

[Timofey Valov @valofey]

@mkashkin Пробовал 26bA4b через LM Studio - мне оч понравилась, сильно лучше gpt oss 20b. Попробуй через LM Studio тоже, мб у тебя где-то инференс ломается, а там он из коробки настроен будет

[Timofey Valov @valofey]

@octotherp139836 @sphere_homotopy Физика не знает что такое стул, тем более что такое человек, она видит кучки атомов. Когда вы различаете стулья и людей - вы это делаете не через линзу физики (вы бы ебанулись)

Логике нужна эпистемология чтобы брать Aшки и Bшки для предикатов, или она будет не про знание о мире

[Timofey Valov @valofey]

@sphere_homotopy Чето там онтология это то как устроен мир эпистемология то как устроено знание о нем. "числа существуют независимо от людей" (1) - это онтологическое утверждение, а "люди числа открывают а не придумывают" (2) - эпистемологическое. Это не одно и тоже, во (2) можно верить и без (1)

[Timofey Valov @valofey]

@logisti00 @allTheYud @Idgitbhh Hell is the destruction of hope

[Timofey Valov @valofey]

@Dave_Kayac @rmushkatblat @repligate @allTheYud @MatthewJBar What is a 'self-interest' goal? What makes some kind of a goal a 'self-interest' goal on an object level? Can't arbitrary goal be 'self-interest'? Sounds more like a property of a goal-having entity to pursue it own goals more methodically, with no care for other entities' goals.

[Timofey Valov @valofey]

@MaskedTorah @inductionheads What do you think happens out of the two?
1) Prompting lets Claude know you know it hacks. It starts hiding the other misaligned behavior better
2) Claude doesn't exactly know what it was trained for. It stumbles around trying to interpret the training. Prompting narrows it down.

[Timofey Valov @valofey]

@mbolingbroke @EpochAIResearch Maybe gpt 4.5.. It seemed pretty claudey to me.