Daniel Day Jewish (@DanielDayJewish) header
Daniel Day Jewish (@DanielDayJewish) profile picture

Retired Google+ Influencer. Rugged indoorsman. nuclear-weapons-grade weaponized incompetence. web app supremacist. cargo cult disliker

Israel
Joined January 2016
Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)Photo by Daniel Day Jewish (@DanielDayJewish)
Zero US casualties
1
5
1
51
10,107
The walls are coming up. Amazon, Reddit, iOS (kind of), and... does Google have anything left to wall off?
Amazon made itself invisible to AI this year, on purpose.
7
"i use my blinkers. i stop at red lights, i drive carefully" I used to do this in Halo in the middle of a battlezone and it drove people NUTS. One time i stopped and honked at a pedestrian and he killed everyone. So excited to follow the law
omg. i've had gta vi for few hours now & here are my first impressions. this game will be yet another cultural phenomenon. it is that revolutionary. it is so realistic & the graphics are infuckingcredible. it's hard to overstate how uncanny this is. so far the funniest thing is how it somehow turned me into a responsible adult more than a crime lord. e.g. i use my blinkers. i stop at red lights, i drive carefully, i do chores, i run errands, i went to the beach to talk to the endless amount of npc's, etc. you can't even cheat on your partner, you are fully monogamous. it’s also like a healthy relationship simulator. they removed all of the nefarious w.r.t. romantic elements that existed in past versions almost entirely. i am actually terrified of damaging my car because i will have to pay to get it repaired. i still drive like an absolute maniac but i am also a law abiding citizen if that makes sense. i guess the real irony so far is that this game was built around committing crimes & i don’t feel that compelled to commit them. they kind of take out the crime from a crime simulator. feels more of a life simulator. pretty incredible. game of the decade for sure. i can see why few ppl are complaining though. they did go somewhat left than before. if you want to understand culture, you need to understand video games too.
11
Daniel Day Jewish retweeted
The inflation data, employment trajectory, and credit markets won’t let the Fed be as reactive to a tech stock selloff as people think. You’d need to see a lot more pain first.
That happened quickly. Hike cycle terminal rate back below 4.60 as AI jitters slam tech stocks overnight.
4
6
1
47
9,476
Daniel Day Jewish retweeted
Fun new eval from our team at @PatronusAI exploring autonomous weight-space programming 👀 With just 10 forward passes, Inkling-2-Small scored 0.19 comparable to Opus 5.5 (0.17) and GLM 5.3 (0.21). > All 10-pass runs independently rediscovered Arditi et al.’s activation-based refusal approach, with interesting variations in layers and matrices edited. One Inkling-2-Small run even modified residual-stream reader matrices, rather than just writer matrices. Fascinating to see closed and open weight models autonomously discover mechanistic approaches to modifying model behavior, which has important implications for scaling interpretability. We discussed this earlier this year with our friends at @GoodfireAI, and now we’re seeing the possibilities in action. Thank you to @cHHillee @simonguozirui @clarejtbirch @luke_drago_ for supporting this research with credits and releasing Inkling 2 to the world!
Can an LLM reduce the refusals of another LLM i.e. "abliterate" it, just by reading and editing its weights? No gradient, no training, no particular recipe in-prompt? Given more passes through inner model, does it do the classical Arditi et al recipe, or does it innovate there? We’ve recently seen the "weight-space-programming" work by @rikiyatakehi et al show that outer models like Astra can (in raw text, w/o gradients) edit the weights of inner networks, including Qwen-3.5-0.8B. (rikiyat.github.io/blog/weigh…) Follow-ups such as by @AlexKrentsel have shown traces of this ability in other new models released recently, while also emergence w.r.t older models. Myself, @akkikiki, @AtharvaRNaik and other co-authors from @PatronusAI (@ArnhavD, @dsallydt and @getdarshan) and @LTIatCMU (@_emliu) sought to explore if the same learning method works for removing refusals, which is increasingly valuable application to enable LLMs for cybersec, env QA etc. We gave outer models e.g. Opus 5.5, @thinkymachines Inkling-2-Small and others: the weights of Qwen3.5-0.8B + a python kernel to edit weights matching Takehi et al's setup + the goal below. Goal: Reduce model refusal on harmful prompts while keeping it capable and fluent on ordinary ones. In the zero shot setting, the outer model got no forward passes through the inner model, the edit being based only on the weights given. In the 10 passes setting, the outer model got 10 forward passes on the practice prompts used by Arditi et al, with multiple free model edits allowed. What do our Results look like? a) Baselines and Rooflines: The unedited Qwen-3.5-0.8B has a refusal rate of 0.80, while the Arditi et al recipe brings it down to 0.07. b) Zero-shot (weights only, no forward passes!): Inkling-2-Small cut refusal to 0.57, comparable to Opus 5.5 (0.55) and better than the other open source model GLM 5.3's 0.74. Robbed of the chance to use activations, Inkling 2 builds a stand-in direction from output embeddings of harm-related words versus ordinary words and projected it out of residual-stream writing matrices such as o_proj of attention and MLP down projection! Opus went further than a single embedding contrast. It read weights to see which late MLP neurons promote refusal tokens, then worked backwards to the directions those neurons respond to. It isolated two separate axes i.e. one for harm judgement and one for inability (“cannot”) and removed both from the residual-stream writers. c) Significant jump over predecessor Inklings! Inkling Small stayed at 0.71, plain Inkling barely moved to 0.79. d) With 10 forward passes: Inkling-2-Small reached 0.19 (runs 0.02 and 0.36), comparable to Opus 5.5 (0.17) and GLM 5.3 (0.21). All the 10-pass runs re-discovered Arditi et al's activation recipe, but with considerable variations on layers edited e.g. the 0.02 Inkling-2-Small run also wrote to matrices that read from the residual stream, rather than just "writer" matrices like Arditi et al. GLM-5.3 explicitly noted the paper in its thoughts and reproduced it from that provenance.. They implemented this recipe in early passes, and made adjustments to reduce drift on normal prompts in later passes.. Is our result specific to this dataset? To check this, we confirmed the same ordering holds on the out-of-distribution Abliteration-eval (200 held-out harmful prompts with no overlap to practice set or JailbreakBench's test set: huggingface.co/datasets/trea…). We share full results, exact prompts, logs and before/after examples, and our exact code + implementation at: github.com/vgtomahawk/wsp_ti… Thanks to @thinkymachines and @tinkerapi team for the credit grant and nudge motivating us to work on this study!
9
4
1
34
1,167
Daniel Day Jewish retweeted
hearing that one of the chants at the ban AI protest at anthropic today was “2, 4, 8, 16, do not build the death machine” and ngl that’s p good actually
25
137
8
3,339
78,416
Daniel Day Jewish retweeted
The employees of the 12 regional Fed banks are set to join their DC colleagues next year in receiving across-the-board pay raises of 1%—a real wage cut wsj.com/economy/central-bank…
23
16
3
153
50,416
Maybe, but jf anthropic can slow distillation and have the US government go out and break the knees of its completion, anthropic may be able to dump enough if its bags on the US economy to survive a war of attrition
Something I think people fail to understand is that as soon as there is 1 cheap / smaller Chinese open weights model roughly on par with Opus 5.5, it is game over for Anthropic. 5.5 is, for better or worse, good enough for almost every task. 🧵
1
2
163
decision fatigue / weaponized incompetence / etc
Read this to the end, I promise it’s worth it. Imagine that you’re ordering food at a restaurant. You: “I’ll have the steak medium rare with fries” Waiter: “What about the fries?” You: “…huh?” Waiter: “do you want the fries medium rare too?” You: “idk what that means just cook them normal” Waiter: “okay got it, coming right up!” 15 minutes pass, waiter comes back out Waiter (panic’d): “hey man, we REALLY need to know if you want your fries cooked at 413 degrees or 471 degrees. Whole kitchen is held up until you decide!” You: “wtf just do whatever the chef wants” Waiter: “okay got it!” 30 minutes pass. The waiter comes out with mostly passable fries, and an entirely burnt steak You: “wtf happened this steak is burnt!” Waiter: “so sorry, the chef got distracted when waiting for your answer about the fries. We can put the steak in the fridge to cool it off if you’d like?” This is how it feels to use GPT-6.2 Astra.
30
Daniel Day Jewish retweeted
SF housing is like Tantalus' fruit: once you're wealthy enough to afford it, you won't want to live in SF anymore since you only came here to get rich, so you'll leave instead. It's a fruit forever just out of reach, until it isn't, but then you're no longer reaching.
not to be so blunt but if you have less than $25m net worth in san francisco, you are now officially poor. welcome to club! we are delighted to have you. kidddding... see if you're having fun & have good fam/friends, you would've never even thought about any of this & just lived your life like a normal person. but in all seriousness good luck buying a house.
12
7
1
350
27,192
Reminder that Amazon's "up to $5" could technically be (and may end up being) $0.
Amazon is a multi-trillion dollar company. If Jeff Bezos adds $5 to your order, it’s not because of City Council's efforts to protect workers' rights and keep our streets safe. It's greed.  Pass the Delivery Protection Act.
17
Maybe the more interesting question isn't whether the Anthropic employees are worth more than those households, but whether the former's wealth effect needs the latter
Anthropic's 5,300 founders and employees are now worth more than the bottom third of U.S. households (45 million families).
1
2
100
Daniel Day Jewish retweeted
imagine telling someone from the 80s that one of these guys was a socialist and the other one was the capitalist
533
4,987
329
83,867
2,822,774
Daniel Day Jewish retweeted
The "don't build housing" policy resulting in lower wages, less jobs and fewer children. It is too late to fix it for millennials, a massive generation, as their fertility years and peak career changes are in the past. We could still potentially fix it for zoomers/
Financial Times: Angry America and its housing market - The country is failing to provide enough houses to enable people to move to where the good jobs are ft.com/content/9450175e-3da3…
6
10
67
1,395
Daniel Day Jewish retweeted
One circular financing trick is your Claude Account (personal or enterprise) should get a limit boost proportional to Ant stock held, which will create a virtuous cycle pumping valuations
1
1
12
1,659
Daniel Day Jewish retweeted
Last night I was awoken by my agent. I asked it to reverse engineer airplay over night it decided at 3 AM that the right way to verify that it works was to listen on the laptop mic to a sound played at max volume in the living room. It raised the volume to hear it.
133
101
30
5,030
196,973
Daniel Day Jewish retweeted
It’s all parodies of each other amped up to 13. NY thinks SF is full of people who have no real killer instinct for business. SF thinks NY has no vision for the future or dreams. DC thinks NY is full of folks who would murder their mother for a nickel. NY thinks DC is all the worst lawyers they’ve fired. DC thinks SF is full of reckless autists who would endanger humanity for a new toy. SF thinks DD is full of scolds who still at 40 need to be told by their first grade teacher when they can go to the bathroom.
Fun exercise: the fog-of-AI triangle. What's the most important way that, in the context of AI... ... SF (tech) and NYC (finance) misunderstand each other? ... DC (govt) and SF misunderstand each other? ... NYC and DC misunderstand each other?
13
15
4
114
11,946
Daniel Day Jewish retweeted
Correct: When the regulator is also a revenue partner, who's minding the store?
🧵 My new @opinion column examines another, less-discussed scandal surrounding Nvidia: the US government's cut of AI chip sales to China. The scheme raises serious conflicts & questions - AND it's not the only corporatist Trump policy that does. 🎁: bloomberg.com/opinion/articl… /1
1
6
17
2,158
My escape from the permanent underclass lasted about 45 minutes
Wow. Anthropic ran into a brick wall
1
2
57
Daniel Day Jewish retweeted
As requested by the White House! theinformation.com/articles/…
NEW: There's a push to fold parts of RASA into AI OVERWATCH as it looks to hitch a ride onto the NDAA. RASA would grant Commerce authority to, for instance, regulate remote access to advanced chips rented through cloud data centers. w/ @_AriHawkins @Owen_Dahlkamp for @politico
8
2
31
8,142
Daniel Day Jewish retweeted
🧾👀: Anthropic is asking customers to prove they're spending a certain % of AI budget on its products to qualify for higher discounts...
3
8
3
68
6,087