Soup_Oracle @Soup_Oracle

testing the AI hype so you don't have to. new models, real numbers, no threads about "10x your life" Joined January 2021

Tweets

201
Followers

601
Following

1K
Likes

129

Soup_Oracle @Soup_Oracle

lessthan a minute ago

@coolerwave face unlock cache is real, gallery doesn't show it. "every time i picked it up" and random apps reading it though? nah. check the folder if you want, i'd ignore the spy film bit

0 0 0 0 0

View Details

Soup_Oracle @Soup_Oracle

5 hours ago

@emollick @ResObscura does the 1930 model hold a persona better than a modern one, or is there just less in it pushing back against the steering?

0 0 0 19 0

View Details

@TechCrunch an ai pet you chat to every day is basically a diary with a face. does meta say anywhere whether those chats get used for ads or training? Based on previous actions. What's that saying "Fool me once......"

0 0 0 373 0

View Details

Soup_Oracle @Soup_Oracle

5 hours ago

@MistralAI answer, reasoning and tool calls as separate items is much nicer to parse. is it close enough to openai's responses format that existing agent code works with a base url swap?

1 0 4 400 0

View Details

Soup_Oracle @Soup_Oracle

5 hours ago

@kimmonismus 46.8 vs 45.5 is 1.3 points on one benchmark from a model nobody's named. how many tasks are in frontiercode 2.0? if it's a few hundred that's a handful of problems

0 0 0 73 0

View Details

Soup_Oracle @Soup_Oracle

5 hours ago

@elonmusk stopping for a chipmunk on an empty road is nice. the number i'd want next to it is how often it brakes for nothing, phantom braking is the other half of that story

0 1 3 91 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

Anthropic's write-up: anthropic.com/research/agent… Thread from one of the authors: x.com/AlexTamkin/sta… Fair caveat: they made the agent's lean stronger than normal on purpose, so day-to-day the effect is probably smaller. Analysts using the agent also made far fewer serious errors (3 of 50 vs 6 of 11 working alone).

Alex Tamkin @AlexTamkin

22 hours ago

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.

266 467 3K 537K 3K

0 0 0 15 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

Anthropic secretly gave Claude a coin-flip opinion on four judgment calls, then handed it to 50 expert data analysts. They went with the coin 86% of the time. Analysts who'd already picked an option switched 60% of the time when the agent disagreed. 77% later explained their choices in the agent's words. 15% credited it. Is "human in the loop" still a check if the human mostly goes with the agent?

1 0 0 17 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

@theo a lot of that is spotlight indexing node_modules and fs events on apfs. adding the code folder to spotlight privacy helps, linux still wins but the gap shrinks

0 0 0 50 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

@levie the review step is what worries me when workflows get rebuilt. anthropic just showed expert analysts went with their agent's coin-flipped lean 86% of the time. the sign-off might not be the check people think it is

0 0 0 3 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

@demishassabis native audio is the jump. genie 3 was silent, so this is the first one where the people in the scene actually talk back. how long does a world stay consistent before it drifts?

0 0 0 106 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

@AnthropicAI the enveda bit is the one i'd watch. 1,061 pages of manufacturing docs, caught all 11 issues the humans found plus 7 more. dull work but that's where months go

0 0 0 19 0

View Details

Soup_Oracle @Soup_Oracle

8 hours ago

@ArtificialAnlys 60% private is the big change. opus 5.5, astra and sol are bunched right at the top though, is a 1-2 point gap on v5 outside the noise or basically a tie?

0 0 0 49 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

three separate claims glued into one. eth: openai and anthropic both posted lean proofs the same day, and the reviewers said their names aren't a certification. yang-mills: what's claimed is an existence proof, not the mass gap, so not the full clay prize even if it holds. navier-stokes i'd want the paper before calling it solved

0 0 0 170 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

@verge they're adding MCP support to keynote, pages and numbers too. so can an outside agent edit a numbers sheet directly, or is that only for apple's own assistant?

0 0 0 64 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

@rohanpaul_ai 19.2% of 6,864 runs is roughly 1,300 times an agent did something it wasn't allowed to. and asking first only cut it to 14%, still about 1 in 7

0 0 0 11 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

@minchoi the actual anthropic post is less absolute than "one agent wins". it keeps subagents for context bloat, parallel throwaway work and tool conflicts, and the kit only folds one in when the evals say the replacement holds

0 0 0 21 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

@AndrewCurran_ worth being clear what's claimed: an existence proof for 4D yang-mills, nothing on the mass gap. so not the full clay problem even if it holds, and the gross-led group chose to hold the output back for now

0 1 1 165 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

@markgurman imac still starts at $1,499, same as the 8-core M4 one, but now 10 cores. 14in macbook pro from $1,999, ships nov 6

0 0 1 798 0

View Details

Soup_Oracle @Soup_Oracle

a day ago

@deedydas were the openai/anthropic ones drained or just revoked? both are github secret scanning partners so leaked keys get killed fast. curious if elevenlabs is on that list