The Washington, DC-based Center for the Study of Organized Hate (CSOH) said it tracked posts that explicitly dehumanised, excluded, and incited violence against Muslims from 1 January to 5 March.

On the day the war began, the volume of such posts surged from just under 2,000 per day to more than 6,000, the report said. "Reposts dramatically amplify the visibility of harmful content, allowing it to spread far beyond the original accounts that generated it."

With reposts included, "the total mention volume of Islamophobic content rises to 279,417, representing an 11-fold amplification of the harmful original posts".

The content examined by CSOH encompassed a wide swath that included everything from personal hate-fuelled opinions to calls for lawmakers to institute strict anti-Muslim policy, including a "Muslim Exclusion Act" and the deportation of all Muslims.

An investigation into the AI bots that appeared on Reddit after they partnered with OpenAI.

Key points:

• The bots post a lot of links to products and services which appear to be adverts but are not marked as such.
• Many of these links are for Sam Altman's World ID.
• Reddit added terms on AI advertising to their business page around the time the bots appeared.
• The bots also make up stories about dead mothers, depression, drug addiction, eating disorders, medical conditions and mental health issues.

https://archive.ph/gx01C

The core thesis of this paper is that the AI community needs to stop treating autonomous agents as just another text generation problem and start building comprehensive infrastructure to support closed loop learning. The authors argue that achieving reliable agentic behavior requires a full stack ecosystem that unifies data synthesis with sandboxed execution and specialized reinforcement learning. To prove this point they introduce the Agentic Learning Ecosystem which consists of an RL framework called ROLL alongside a sandbox manager named ROCK and an agent interface known as iFlow CLI. They believe that isolating models in static training environments is a dead end for solving complex real world workflows.

The team developed an open source model named ROME using a tightly integrated training pipeline with reproducible execution environments which allowed a relatively small 30 billion parameter model to rival or beat massive proprietary models exceeding 100 billion parameters on difficult software engineering benchmarks.

A big part of their argument rests on the idea that credit assignment in reinforcement learning needs to change. They propose a novel algorithm called Interaction Perceptive Agentic Policy Optimization which shifts the reward focus from individual text tokens to broader semantic interaction chunks. This chunk level optimization stabilizes the training process over long horizons and prevents the policy collapse often seen in complex tool use scenarios.

We're increasingly seeing a shift of priorities away from raw data scale and focus on the systematic infrastructure as the actual bedrock of next generation models.

cross-posted from: https://lemmy.ml/post/44059967

for those not familiar with Mark Pilgrim, he is/was a prolific author, blogger, and hacker who abruptly disappeared from the internet in 2011.

cross-posted from: https://lemmy.bestiver.se/post/968527

HN comments

"Breaking Free: Pathways to a fair technological future" is a new report from Forbrukerrådet. The report itself is a light read: it's in English, and while it is 100 pages long [PDF], it is in fact enjoyable and even amusing – we laughed quite a few times when reading it. For one thing, it contains a surprising number of puns and the occasional starred-out swearword, such as "Do androids dream of electric s***." A stodgy bureaucratic report this is not.

https://youtu.be/T4Upf_B9RLQ

I had been using a magic eraser (never on the screen) and have not yet suffered any ill effects as a result (although I also rarely clean my laptop at all), but I gather that this is not recommended any longer (if it ever was). Alcohol wipes are good for the screen, but not as effective for the keyboard and other non-screen parts. Any suggestions?

An Amazon Web Services data center in the United Arab Emirates suffered a multi-hour outage on Sunday after unidentified “objects” struck the facility and triggered a fire. The incident occurred around 4:30 a.m. local time and affected the availability zone mec1-az2 in the ME-CENTRAL-1 region.

The fire department cut power to combat the flames, resulting in significant disruptions to cloud services. Given the simultaneous Iranian retaliatory attacks on the Gulf states, suspicion arises that the impacting objects may have been missiles or drones. Amazon has not confirmed anything on its part.