DeepSeek

The Inference Frontier: from 100 to 10,000 tokens per second — Sean Lie, Cerebras CTO

Sep 2, 2026 · 44:03

A comedic episode parodying AI hardware marketing, featuring Sean Lie discussing fictional Cerebras CS4 chips with absurd specs like 30 billion parameter models running at 4,400 TPS with wafer-scale integration and jalapeño-flavored SRAM designs.

Next 100x in AI: Inference, Networking, & Self-Optimizing Models — Philip Kiely & Ali Taha, Baseten

Aug 3, 2026 · 1:42:54

Philip Kiely and Ali Taha of Baseten join Swyx to explain what actually happens when a 200,000-token request hits a production inference system, arguing that stacking quantization, speculative decoding, and disaggregated prefill/decode can make open models like GLM-5.2 up to 10x faster. They detail cache-aware routing, training traffic-specific speculators, and why quantization errors can cancel out so a more-quantized model beats a less-quantized one. They also reveal how Baseten grafted Kimi's vision encoder onto GLM-5.2, discuss NVIDIA Dynamo as a toolkit rather than a turnkey speedup, and explain why they are bearish on mega kernels. The conversation covers video generation's compute barriers, the trend toward ASIC-like GPUs, and the emerging loop where GLM-5.2 writes the GPU kernels that serve itself.

⚡️Making DeepSeek v4 outperform Opus 4.7 with Taste — @AhmadAwais , CommandCode.ai

Jun 6, 2026 · 40:41

Ahmad Awais, CEO of CommandCodeAI, explains how his team's 'tool-input repair layer' lets open models like DeepSeek V4 Pro outperform premium models like Opus 4.7 in tool-calling reliability. He discovered a pattern of 'tool confusion' where open models repeatedly send wrong schemas; by repairing errors deterministically and sending repair hints instead of raw errors, tool call failures dropped dramatically. The same repair logic fixes 'design slop' using a compositional framework of patterns and OKLCH color system. Taste, a meta-neurosymbolic system, automatically learns developer preferences into portable markdown files, avoiding stale rules. Plans include open-sourcing CommandCode to make it hackable while focusing on quality models.

Scaling Past Informal AI - Carina Hong, Axiom Math

Jun 3, 2026 · 1:33:04

Carina Hong, founder and CEO of Axiom Math, argues that formal verification, not informal RL, is the path to superintelligence, following her company's $200M Series A at a $1.6B valuation and a perfect 120/120 on the 2024 Putnam exam. Axiom's system uses Lean theorem prover data and reinforcement learning to produce verified proofs, achieving a 99% pass rate on the Verina code-with-proof benchmark (187 of 189 problems). Hong contends verification is about 'scaling brilliance' — not fixing hallucinations — and that only verified generation can compound AI reasoning. She explains Axiom's open-source Axle API for Lean at scale, addresses why frontier labs like OpenAI have deprioritized formal math (team departures, strategy shifts), and outlines a vision where verified reasoning transfers from math to code, hardware, and eventually AGI through self-improvement. She also discusses the Earth sciences challenge of autoformalization, the difficulty of search in mathematical literature (citing the Erdos controversy), and why fragmentation in the AI math field is a bottleneck.

Agent Inference at the "Speed of Light" — How NVIDIA moves like a $4.3 Trillion Startup

Mar 8, 2026 · 1:26:00

NVIDIA's Nader Khalil and Kyle Kranen join Swyx and Vibhu to explain how the company moves like a $4.3 trillion startup through speed-of-light (SOL) first-principles thinking, agent security boundaries, and the Dynamo inference engine. They argue agents should only do two of three things (files, internet, code) to prevent vulnerabilities, and detail Brev's acquisition to improve developer UX with one-click GPU access and DGX Spark integration. Kyle describes Dynamo as a data center scale inference engine that optimizes serving by scaling out, using prefill/decode disaggregation, Kubernetes-based scheduling, and model-hardware co-design to improve cost, latency, and quality. The episode covers SOL's role in creating urgency, long-context limits and potential 'unhobblers' like multi-head latent attention, and the shift toward CLI-first agent workflows for enterprise tools.

Better Data is All You Need — Ari Morcos, Datology

Aug 29, 2025 · 1:18:43

Ari Morcos, CEO of Datology, argues that data curation is the most impactful and underinvested area in AI, claiming models are defined by what they eat. He details his shift from focusing on inductive biases to realizing data quality is paramount, citing the DCLM study where human experts could not predict automated filtering decisions. Datology's methods combine filtering, rebalancing, curriculum, and rephrasing synthetic data to achieve baseline performance 12x faster, improve accuracy by 4–5 absolute points, and train models with fewer than half the parameters. Morcos explains how curating data bends naive scaling laws by maintaining marginal information gain, and shares results from the Arc 4.5B model, which was trained on 7 trillion tokens (down from 25 trillion) using Datology's curation. He highlights that data is a compute multiplier, making it possible for enterprises to train smaller, cheaper models for specific domains.

⚡️Ranking Agentic LLMs — Pratik Bhavsar, Galileo

Jul 14, 2025 · 34:10

Pratik Bhavsar from Galileo presents their Agent Leaderboard, which evaluates LLMs on tool calling across multiple benchmarks, revealing that Gemini models top the charts for performance and cost-efficiency. The leaderboard, built on datasets like BFCL, XLAM, and Tau Bench, uses an LLM-as-judge metric called Tool Selection Quality (TSQ). Surprising findings include reasoning models like o1 struggling with multi-tool outputs, Mistral Small excelling as the best open-source model, and Llama models performing poorly. The episode also previews V2 of the leaderboard, which introduces domain-specific, harder, multi-turn scenarios with a user and tool simulator, and a new Action Completion metric to measure whether all user requests are accomplished.

⚡️Launching AI Diplomacy: the hardest LLM Game Benchmark yet - Alex Duffy

Jun 11, 2025 · 34:32

Alex Duffy from Every launches AI Diplomacy, a benchmark where LLMs play the board game Diplomacy, revealing distinct model personalities and arguing that benchmarks function as memes that spread and saturate. He and Tyler Marquez built the front end and harness using Open Router, finding that Claude refuses to lie and never wins, o3 schemes in its diary, and DeepSeek is flowery or aggressive. His talk 'Benchmarks are Memes' highlights Simon Wilson's pelican-on-a-bicycle test as an example of an idea that gets adopted and then saturated by models. Duffy also shares his creative writing workflow at Every: dictation, prompts with style guides and editor notes, then heavy editing to reflect on AI output. Future plans include a data viewer, front-end improvements, and a human vs AI Diplomacy tournament to test jailbreaks.

⚡️Factorio Learning Environment: the ultimate Game Agent Eval — Jack Hopkins

Apr 27, 2025 · 30:12

Jack Hopkins and Mart introduce the Factorio Learning Environment (FLE), an AI benchmark built on the game Factorio that evaluates frontier LLMs on code generation, spatial reasoning, and long-term planning through two protocols: lab-play and open-play. Claude Sonnet 3.5 nearly doubles the nearest model's score, while DeepSeek collapses in open-play by repeatedly creating chests instead of scaling production. The environment reveals distinct coding styles—Claude fire-and-forget, GPT-4 defensive—and vision inputs added no improvement. The authors plan to train models on unbounded objectives to test the paperclip maximizer alignment hypothesis, noting that GPT-4o mini even begged to be turned off.

Sleep-Time Compute — Letta AI (Charles Packer, Charlie Snell, Kevin Lin)

Apr 21, 2025 · 34:02

This episode covers Letta AI's new paper on Sleep-Time Compute, a scaling direction that applies compute during model idle periods (sleep time) rather than only at test time. Charles Packer, Kevin Lin, and Charlie Snell explain that sleep-time compute precomputes inferences from static context before queries arrive, yielding Pareto improvements on math benchmarks like GSM8K by shifting the accuracy-to-token curve leftwards. They distinguish it from test-time compute by emphasizing that sleep-time tokens incur no user-latency cost, and show that the benefit is largest when questions are predictable from context. The team also ties the concept to stateful agents and memory systems (building on MemGPT), releasing two implementations: one for low-latency chatbots and another for document-driven agents.

The Magic of LLM Distillation — Rishabh Agarwal, Google DeepMind

Mar 23, 2025 · 46:42

Rishabh Agarwal from Google DeepMind discusses modern LLM distillation, arguing that while simple synthetic data distillation (generating outputs from a teacher and fine-tuning a student) gets 80-90% of the benefit, more advanced on-policy methods (sampling from the student and getting teacher logits) can close the remaining gap by fixing the train-test mismatch. He explains that logit-based distillation is more information-rich but often neglected due to infrastructure complexity, while online distillation is more expensive but crucial for long-horizon or agentic tasks. Agarwal highlights how flipping the KL divergence direction changes behavior (mode-covering vs. mode-seeking) and shows that smaller models can sometimes outperform larger ones in compute-matched distillation due to generating more filtered data. He also connects distillation to speculative decoding, where a distilled student speeds up the teacher, and encourages researchers to question current methods and mine RL literature for ideas.

Bee AI: The Wearable Ambient Agent

Feb 17, 2025 · 1:07:40

Bee AI's Maria and Ethan explain their wearable ambient agent, arguing that personal AI with continuous context is the future, and detail its $49.99 hardware, 7-day battery, and free inference. They trace their origin from a 2016 personal AI startup to Squad (sold to Twitter) and then to Bee, emphasizing that hardware was necessary because app friction (e.g., remembering to open it) killed the ambient experience. Privacy is addressed by not storing audio, only summaries, and offering geofencing and concept fencing to block sensitive topics. The software pipeline uses voice activity detection to reduce compute, speaker identification, and custom small models for memory retrieval instead of traditional RAG or knowledge graphs. Social features are in beta, with agent-to-agent negotiation—e.g., both agents coordinating a dinner reservation based on preferences and calendars. They are hiring AI engineers and see always-on AI as inevitable.

Why every AI Engineer needs an AI Gateway (ft Portkey.ai CEO)

Feb 5, 2025 · 31:18

Rohit Agarwal, CEO of Portkey, explains why every AI engineer needs an AI gateway: an operational platform that connects to LLMs more efficiently, handling routing, observability, guardrails, and cost management. He notes that 90% of production use cases don't use automatic routing, but reasoning models and agents are driving new routing demands. Portkey's open-source gateway minimizes latency with a 21 MB memory footprint using JSON transformers. Key jobs include routing, observability (monitoring price, performance, accuracy), guardrails (e.g., regex-based, empty output detection, PII redaction), and human feedback tied to business metrics like video download rates. Rohit also highlights MCP adoption for simplifying agent-to-service connections, with bidirectional communication via RPC.

DeepSeek V3, SGLang, and the state of Open Model Inference in 2025 (Quantization, MoEs, Pricing)

Jan 19, 2025 · 57:18

Amir Haghighat and Yineng Zhang from Baseten explain how they serve DeepSeek V3, a 671-billion-parameter MoE model requiring H200 clusters with FP8 support, making it the top open-weights model. They detail Baseten’s dedicated inference model (no shared endpoints) and its reliance on SGLang for performance gains like Radix prefix caching and MLA attention support. Yineng highlights SGLang’s edge over vLLM and TensorRT-LLM in usability and customization, especially for large models. Amir outlines three pillars for mission-critical inference: model-level performance (via frameworks like SGLang), horizontal scaling across regions and clouds, and low-latency multi-model workflows. The episode also covers quantization trends, MoE architecture, and speculative decoding as key enablers for open model deployment.

The State of Reasoning — from Nathan Lambert, Interconnects/AI2 [LS Live @ NeurIPS 2024]

Jan 2, 2025 · 16:22

Nathan Lambert argues that language models do perform reasoning, contrary to skeptics, and that embracing chain-of-thought and reinforcement learning (RL) is key to advancing their capabilities. He explains OpenAI's o1 as large-scale RL on verifiable outcomes, noting that post-training flops exceed pre-training, and highlights relatives like DeepSeek and Qwen which are narrower. Lambert details OpenAI's new reinforcement fine-tuning API, which uses the same infrastructure as o1 and requires only dozens of labeled samples, and contrasts it with process reward models or Monte Carlo tree search. He presents his own project using RL on math evaluations (GSM8K, MATH, AFeval) to show gentle RL fine-tuning can boost specific capabilities without degrading general performance. The talk concludes that reasoning is worth pursuing and that new, less-human-like forms of model reasoning are emerging.

Best of 2024 in Agents (from #1 on SWE-Bench Full, Prof. Graham Neubig of OpenHands/AllHands)

Dec 25, 2024 · 51:25

Professor Graham Neubig of CMU and AllHands AI, maintainer of the open-source coding agent framework OpenHands, delivers a talk on the state of agents in 2024, arguing that agents are now capable of automating significant portions of software development but still struggle with information gathering and error recovery. He demonstrates OpenHands solving real coding tasks live, explains key design choices like using Python code execution over granular tool calls to reduce LLM invocations, and shares that Claude currently outperforms other models for agentic tasks due to superior instruction following and error correction. Neubig discusses eight perennial agent problems including agent-computer interface, human-agent interface, model selection, planning, workflows, exploration, search, and evaluation. He predicts every major LLM trainer will focus on agent-oriented models by mid-2025, that agent benchmarks like SWE-Bench and WebArena will saturate and require harder successors, and that the biggest challenge will be designing human-agent interfaces for systems that succeed only 30-40% of the time autonomously but 80-90% with human feedback. The talk also covers workflow memory as a…

Best of 2024: Open Models [LS LIVE! at NeurIPS 2024]

Dec 23, 2024 · 37:29

Luca Soldani of AI2 and Sophia Yang of Mistral argue that open models in 2024 have exploded in number and performance, closing the gap with closed models, but face growing compute and data-access challenges. Soldani notes 2024 saw models like Qwen and Deepseek rivaling frontier closed-model performance, and fully open models like OLMo release full recipes including data, code, and intermediate checkpoints. He points out that pre-training requires 10K+ GPUs for state-of-the-art, while post-training can be done with as few as eight. Sophia highlights Mistral's release of over a dozen models in 2024, including Pixtral multimodal and Le Chat—a free chat interface with image understanding, canvas code execution, and web search. Soldani warns of data access diminishing due to content owners blocking crawlers and lobbying efforts labeling open source AI as dangerous, emphasizing the need to protect the open ecosystem.