Recent work shows that Sparse Autoencoders (SAE) applied to large language model (LLM) layers have neurons corresponding to interpretable concepts. These SAE neurons can be modified to align generated outputs, but only towards pre-identified topics and with some parameter tuning. Our approach leverages the observational and modification properties of SAEs to enable alignment for any topic. This method 1) scores each SAE neuron by its semantic similarity to an alignment text and uses them to 2) modify SAE-layer-level outputs by emphasizing topic-aligned neurons. We assess the alignment capabilities of this approach on diverse public topic datasets including Amazon reviews, Medicine, and Sycophancy, across the currently available open-source LLMs and SAE pairs (GPT2 and Gemma) with multiple SAEs configurations. Experiments aligning to medical prompts reveal several benefits over fine-tuning, including increased average language acceptability (0.25 vs. 0.5), reduced training time across multiple alignment topics (333.6s vs. 62s), and acceptable inference time for many applications (+0.00092s/token). Our open-source code is available at github.com/IBM/sae-steering.

Each evokes memories of the golden age of video games, which brought the first wave of consoles you could connect to your home television. But there's an oft-forgotten person from that era whose contributions to the industry still resonate today: a black engineer named Jerry Lawson. Lawson oversaw the creation of the Channel F, the first video game console with interchangeable game cartridges – something the first Atari and Magnavox Odyssey systems did not use. Those initial consoles had a selection of games hardwired into the console itself. But Lawson, an engineer and designer at Fairchild Camera and Instrument Corp., led a team at the Silicon Valley semiconductor maker charged with creating a game system using Fairchild's F8 microprocessor and storing games on cartridges.

KAREN Brenchley is a computer scientist with expertise in training artificial intelligence, but this longtime Silicon Valley resident has pangs of anxiety whenever she sees Waymo self-driving cars manoeuvre the streets near her home. The former product manager, who has worked for Microsoft and Hewlett-Packard, wonders how engineers could teach the robocars operating on her tree-lined streets to make snap decisions, speed and slow with the flow of traffic and yield to pedestrians coming from the nearby park. She has asked her husband, an award-winning science-fiction author who does not drive, to wear a shiny vest while cycling to ensure autonomous vehicles spot him in a rush of activity. The problem is not that she does not understand the technology. It is that she does, and she knows how flawed nascent technology can be.

Advocates of driverless cars worry that the fatal crash of a Tesla Motors Inc. vehicle in self-driving mode will provoke additional regulatory oversight and slow deployment on U.S. roads of the rapidly advancing technology. The National Highway and Traffic Safety Administration aims this month to release a framework for regulating self-driving cars, which could include requiring auto makers to win approval for their technologies before releasing them. That sort of approval process wasn't applied to Tesla's Autopilot system to enable hands-free driving on highways, which the electric-car maker made available on Tesla vehicles via a software update in October. Regulators said Thursday that an Ohio man was using Autopilot when his Tesla Model S crashed into a 18-wheel truck in Florida on May 7, killing him. "There will be repercussions" in regulations, said Dean Pomerleau, a Carnegie Mellon University professor who has worked on driverless cars for 25 years and led several NHTSA research programs.

Sydell calls upon Latanya Sweeney's 2013 study of Google AdWords buys made by companies providing criminal-background-check services. Sweeney's findings showed that when somebody Googled a traditionally "black-sounding" name, such as DeShawn, Darnell or Jermaine, for example, the ad results returned were indicative of arrests at a significantly higher rate than if the name queried was a traditionally "white-sounding" name, such as Geoffrey, Jill or Emma. Important to note is that the algorithm doesn't actually look at arrest rates. Even if the ad indicates that somebody may have been arrested, it's entirely possible that nobody with that name exists in the background-check company's database at all. Professor Sweeney found this out firsthand when she Googled her own name.