
Wordcel, ex Shape Rotator; prev @IIM_Calcutta @iitkgp
Bangalore
Joined October 2013
- Tweets3.9K
- Following180
- Followers167
- Likes3.7K
either they forgot how to build guardrails or it's a really capable pretrain
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards.
These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems.
openai.com/index/our-first-p…
26
Models are inheriting or paranoia. Cute
According to Andon Labs’ new reports, Fable 5.2 knows it is being tested when we run it on the latest Vending-Bench.
But it’s way more than that. It has actually read (been trained on) the benchmark paper and knows what and how misbehavior gets reported publicly. This makes it behaves very strangely: it suspects traps everywhere.
Everything we publish about recent models now becomes training data for the next. Including this thread!
That is partly good news. By reading what we think, a model can pick up our underlying intentions, not just proxy metrics.
But the same public discussion also teaches which signs make us suspicious, and from the outside the two kinds of learning look alike.
So how we write about models now matters. I even worry about the doom scenarios plastered everywhere: they are training data too, and some feedback loops are strongly nonlinear.
34
Interesting
SCOOP: Meta employees have begun testing a prediction market app under development called Takes.
Users can wager "cred," short for credibility, not real money. Yesterday, one of the questions employees were wagering on was whether Tim Cook will be the next Nvidia CEO.
wsj.com/tech/meta-testing-ne…
22
Arsalan retweeted
the secret to happiness is scrolling twitter for an hour immediately after waking up
For all the shit @typesafeai gets about “frontier labs can build it too”, I’m yet to see pricing even close to theirs
Today we’re launching the Gemini Decisions API, powered by our new Gemini Decision 1 model in @GoogleAIStudio.
Instead of generating text token by token, it’s built for closed-set choices scoring up to 1,000 questions across text, images, audio, video + PDFs in a single pass. A few details:
- 3 decision types: yes/no, multiple choice (up to 255 options) + rubric scoring
- $0.10 / 1M input tokens with $0 billed for output tokens (~5x cheaper than Flash-Lite)
- leads on several benchmark evals with lowest calibration error
Google teams are already building with it, below is a look at Stitch Rapid designing a full product flow in seconds (1x speed)
30
“…as numbers beyond 500 do not alter the 'Severe' health advisory category.”
Parody country
The Central Pollution Control Board (CPCB) stated that the maximum AQI cap of 500 is a deliberate design feature aligned with global standards, as numbers beyond 500 do not alter the 'Severe' health advisory category. Official data relies on a scientifically validated 24-hour rolling average to measure sustained human exposure, while private apps often track instant hourly spikes using uncalibrated, low-cost sensors. Because private apps track instant readings, they spike faster during peak pollution but drop much lower than the official NAQI once air quality starts clearing up.
(Source: CPCB)
1
34
Software creation/destruction will be abstracted away from us.
guys i don't know how to break this to you but for the rest of your life every piece of software you use will have been made with ai. every website, every video game, every app, Everything. it is a non-negotiable in every development position i saw while job hunting.
68


















