Claude 3.7 Sonnet from Anthropic took on Pokémon Red with no prior training, using only vision, memory, and function calls. With its extended thinking mode, it reasoned through challenges, defeating 3 gym leaders by dynamically adapting to the game’s mechanics. This is.

DeepSeek-R1 discovered a powerful compute multiplier for LLM reasoning capabilities. Instead of scaling through brute-force compute or complex architectures, they found an elegant shortcut. The Multiplier (GRPO):.• Generate multiple solutions & let them compete.• Learn from.

too much computing is like too much money. they both stifle innovation.

RT @JeffBezos:

2024 Belongs to LLM-Based Coding Agents and Why We Still Need Time-Honored Coding Best Practices (For Now). Of all the remarkable AI breakthroughs emerging this year, LLM-based coding agents have been turning the most heads. Their ability to generate workable code on the fly is.

Transformers: The Surprising One-Stop Shop for Deep Learning and AI. People have been wary of “all-in-one” solutions in computing for ages—especially in distributed systems, where a single architecture can seem riskier than more specialized approaches. As Tanenbaum and van Steen.

Some Thoughts on Scaling Model Reasoning with Test Time Compute. How models determine compute allocation at test time and what “difficulty” really means for LLMs?. Determining how much compute(time) to apply at inference is probably far from exact. Models may gauge complexity by.

One of the best perks of driving an EV? Guilt-free idling. Imagine sitting in your car, waiting for your kid’s math class to end—no emissions, no engine noise, and no wasted fuel. You can keep the air on, charge your phone, or even catch up on some podcasts, all while being.

I’ve been thinking about how current AI coding assistants are pretty impressive. They handle code completion, offer contextual suggestions, optimize algorithms, and can even help write small-scale applications with a bit of tinkering. I think, the next logical step from here.

Influence Without Authority is So Tough—But So Worth It When You Nail It. Trying to get things done without a fancy title or authority is tough. You can’t just tell people what to do. Instead, you’ve got to get them on board by earning their trust, proving your value, and showing.

comfort is an addiction.

Can We Build Deterministic Systems on Top of Probabilistic Foundations Like LLMs?. It's a fascinating question that we don't have a definitive answer to yet. However, we can draw an intriguing parallel from the world of physics. In quantum mechanics, the microscopic world is.

My speculative take—based on publicly available information—on how the recently released OpenAI o1 model advances the state of the art (SOTA) in reasoning tasks is that it combines existing techniques in a new, scalable recipe. A key shift is moving significant compute from.

this one tiny paragraph explains what 'inference' is really about better than entire volumes of literature on ML.

perfection is not a computable operation.

As Builders, Should We Wait for LLMs to Become Deterministic? Lessons from Computing. Unpredictability, often seen as a limitation, can be a powerful tool in computing:. 1/ Evolutionary Algorithms use randomness through mutations and crossovers to explore vast search spaces,.

The Long History of Tech’s Unintended Side Effects: Serendipity or Systematic Design?. Tech history is full of unexpected side effects that revolutionized fields beyond their original purpose. GPUs, for example, were designed to solve a specific class of problems very efficiently.

Why LLM Agents Have Taken Off in Coding Domains but Not in Others - A Thought. LLM-based coding agents have gained significant traction, particularly because the code they generate can be easily verified and refined using existing tools. This success is largely due to the fact.

Computer geeks, what's one accessory under $100 that you absolutely love?.

Today, OpenAI announced GPT-4o-mini: 15 cents per million input tokens, 60 cents per million output tokens, MMLU of 82%, and fast real-time responses. The trend of rapidly reducing inference costs continues and is gaining momentum. These smaller but highly capable models will.