AI Weekly Roundup Feb 8–14 2022: InstructGPT, Waymo SF Expansion, EU AI Act, and the Chip Crunch
TL;DR
OpenAI published early findings on InstructGPT: a 1.3B parameter model that humans prefer over vanilla GPT-3 (175B) thanks to RLHF. Waymo secured a California deployment permit to expand its autonomous vehicle service across San Francisco. EU Parliament committees started marking up the AI Act with draft opinions landing this week. OpenAI is reportedly expanding access to DALL-E for more researchers and creatives. And the global semiconductor shortage continues to hammer GPU supply for AI training workloads.
OpenAI's InstructGPT: Small Model, Big Alignment Win
OpenAI dropped research this month on InstructGPT, and if you've been following the alignment discourse, this one matters. The core result: a 1.3 billion parameter model fine-tuned with reinforcement learning from human feedback (RLHF) is preferred by human evaluators over the 175B parameter GPT-3, a model literally 100x its size.
The training pipeline is a three-stage process. First, supervised fine-tuning on human-written demonstrations. Second, training a reward model on comparison data, where humans rank multiple outputs and the model learns what "good" looks like. Third, using proximal policy optimization (PPO) to fine-tune against that reward model.
The implications are significant. InstructGPT outputs are preferred over GPT-3 85% of the time in head-to-head evaluations. It generates less toxic output, fewer hallucinations, and actually follows instructions instead of pattern-completing its way through your prompt. This is OpenAI publicly stating that scaling model size alone isn't the path to useful AI. Alignment technique matters more than parameter count.
For you homelab runners and indie hackers thinking about local model deployment: pay attention to this research. If a 1.3B model can beat a 175B model on real-world usefulness, the future of practical AI might fit on hardware you can actually afford.
Waymo Gets Green Light to Expand Driverless Service in San Francisco
Waymo picked up a CPUC deployment permit in California this month, clearing the path to expand its driverless taxi operations across San Francisco. The company's Trusted Tester program, which kicked off in August 2021 with its all-electric Jaguar I-PACE fleet, is now operating under formal regulatory approval.
Here's the competitive landscape that makes this interesting. Cruise (GM's autonomous driving subsidiary) is also running driverless rides in SF, and the two companies are effectively in a land-grab for regulatory permits and rider trust in the same city. Waymo's fleet still rolls with a safety operator in the front seat for its Trusted Tester rides, but the regulatory framework is clearly moving toward fully driverless commercial service.
The permit authorizes 24/7 operations and fare-charging capabilities. If you're building in the autonomy space or just watching the self-driving timeline, San Francisco is the proving ground. Two companies, same streets, competing approaches. The regulatory posture from California suggests the state is betting on managed expansion over restrictive gatekeeping.
EU Parliament Committees Begin Shaping the AI Act
The EU AI Act entered a critical phase this week as parliamentary committees started publishing draft opinions. The Special Committee on Artificial Intelligence in a Digital Age (AIDA) dropped its draft opinion on February 11, and the CULT committee's opinion followed shortly after on February 21.
Quick context if you haven't been tracking this: the European Commission proposed the AI Act in April 2021. Lead negotiators Brando Benifei and Dragoş Tudorache were confirmed in December 2021, and the internal market and civil liberties committees are jointly leading the process. This week's activity represents the first substantive markup of the legislation.
What should you care about? The Act introduces a risk-based classification system for AI applications. High-risk systems (think biometric identification, critical infrastructure, employment decisions) face strict requirements around transparency, human oversight, and data governance. If you're shipping AI products to European users (or plan to), the compliance requirements being debated right now will define your obligations.
The Commission also presented a new Standardisation Strategy on February 2, signaling that technical standards for AI systems are being developed in parallel with the legislation. This isn't just a policy document. It's the framework that will determine how "compliant AI" gets defined in practice.
DALL-E Access Continues to Expand
OpenAI is widening the aperture on DALL-E access. Since revealing the text-to-image model in January 2021, access has been tightly controlled through a research preview available only to vetted participants. Reports indicate the company is gradually expanding that pool, giving more researchers, artists, and developers hands-on time with the system.
The expansion signals OpenAI's growing confidence in its content safety mechanisms. The controlled rollout has always been about managing risk: DALL-E can generate photorealistic imagery from text prompts, and the potential for misuse is obvious. By widening access incrementally, OpenAI is collecting usage data and stress-testing its filters before any broader public release.
For the broader AI community, DALL-E remains the most visible demonstration of multimodal AI capabilities. The model combines language understanding (via a modified GPT-3 architecture) with image generation, and the results are genuinely impressive. If you've been on the waitlist, your odds of getting in just improved. If you haven't signed up, now might be the time, since momentum suggests a broader rollout is coming.
The GPU Shortage Keeps Squeezing AI Labs
The global semiconductor shortage, now entering its second year, continues to create real bottlenecks for AI training workloads. NVIDIA GPUs remain scarce, and if you've tried to buy an A100 (or even a consumer RTX 3090 for homelab experiments), you already know the pain.
The root causes haven't changed. COVID-disrupted supply chains collided with surging demand from crypto mining, remote work infrastructure, and accelerating AI research. TSMC and Samsung are running near capacity on the advanced nodes (7nm, 5nm) that produce the chips powering modern AI hardware. Lead times for high-end components are stretching past 40 weeks.
This shortage has real consequences for the AI research landscape. Well-funded labs at Google, Meta, and OpenAI can pre-order thousands of GPUs. Startups, academic researchers, and indie hackers get the scraps. The America COMPETES Act, currently moving through Congress, proposes $52 billion in domestic semiconductor manufacturing investment, but new fab capacity takes years to come online.
If you're running a homelab or bootstrapping an AI startup, your options right now are cloud compute (expensive), used hardware (limited), or patience (painful). The shortage is shaping who gets to do AI research and who doesn't, and that's a problem that won't resolve in 2022.
What Else Crossed the Wire
- Google Research published updates on scaling language models with Mixture-of-Experts architectures, continuing the trend of efficiency-focused ML research.
- Meta AI (formerly Facebook AI Research) continued open-sourcing research tools, maintaining its position as the most open of the big labs.
- The America COMPETES Act advanced in Congress with bipartisan support for semiconductor funding, though the final bill's shape remains uncertain.
- Hugging Face keeps growing as the de facto hub for open-source model hosting and collaboration, with new model uploads accelerating week over week.
Key Takeaways
- RLHF is the alignment technique to watch. InstructGPT proves that a small, well-aligned model beats a large unaligned one. This reshapes the compute-vs-technique debate.
- San Francisco is the autonomous vehicle battleground. Waymo and Cruise are both operating in the same city with regulatory blessing. The next 12 months will be decisive.
- The EU AI Act is being written now. If you ship AI products to Europe, the compliance framework being debated this month will affect you directly. Start paying attention.
- Multimodal AI access is expanding. DALL-E's widening availability signals that text-to-image models are moving from research curiosity toward production readiness.
- Hardware access determines who builds AI. The chip shortage isn't just an inconvenience. It's a structural barrier that favors incumbents over startups and hobbyists. Cloud costs and used GPU markets are your near-term workarounds.