The Daily FM — a daily podcast built from the sources you actually read

Subscribe to this podcast

AI Daily August 9: Anthropic Claude Code Adds Cross-Machine Agent Handoffs as DeepSeek Prices RiseHere is today's AI Daily for Sunday August 9th. Yesterday, Anthropic expanded Claude Code with session-to-session messaging, allowing one coding session to hand a summary to another session on a different machine without transferring the entire working history. That sounds modest, but it is an important building block for multi-agent development. Instead of one giant, fragile context window, teams can divide work among specialized agents and preserve just the relevant handoff. Anthropic is also moving toward classifier-mediated automation as the default, making reliability and permissions as central as the model itself. A second story is the economics of AI coding. DeepSeek users spotted an in-product notice saying its API prices will rise significantly soon, though the company has not yet published final rates or an effective date. DeepSeek has been a favorite for low-cost, high-volume coding and agent workflows. If the increase is substantial, developers may shift traffic to other hosts serving the same open models, or put more emphasis on routing requests among models by cost, speed, and availability. Meanwhile, AI video is becoming more usable outside lab demos. Seedance 2.5, ByteDance’s latest video model, is rolling out through creative tools including Photo AI. Early user demonstrations yesterday showed generated presenter-style clips using uploaded images and a voice sample. These examples are not independent benchmarks, and visible mistakes remain, but the direction is clear: convincing short-form synthetic video, complete with speech, is becoming accessible to ordinary creators in minutes. And early today, the “Kill My SaaS” coding-agent competition released its first LLM-as-judge evaluations. More than 600 people applied, with entrants using any coding agent and up to 500 dollars in token spending to replace costly business software. It is a small event, but it captures a growing shift: AI coding is being judged less by isolated benchmark tasks and more by whether it can produce a working replacement for a real product. The broader trend is that agent progress is moving into workflows, economics, and evaluation. The strongest model alone is not enough; teams increasingly need safe handoffs, effective harnesses, budget controls, and proof that the resulting software actually works. Thank you for listening to AI Daily from The Daily FM. See you tomorrow!

Subscribe to this podcast

Financial Markets August 9: Hormuz Stalemate, Aramco Fire and Moore Threads’ Hong Kong ListingHere is today's Financial Markets for Sunday August 9th. Markets are closed for the weekend, so news is likely to be lighter, but investors will return tomorrow with energy risk and global growth signals firmly in focus. The Strait of Hormuz remains the market’s largest immediate uncertainty. Bloomberg reports that a reopening remains elusive, with Iran insisting that its terms be met before an agreement can move forward. That keeps a geopolitical risk premium embedded in oil and shipping markets. Early today, a fire also broke out at Saudi Aramco’s Jazan refinery, though Saudi Arabia’s energy ministry said the blaze was extinguished. The incident adds to the sensitivity around regional energy infrastructure, even without a reported indication of prolonged disruption. China offered a potentially helpful inflation signal. Bloomberg says factory-gate inflation eased for the first time since the Iran war began in late February, while consumer-price growth also slowed. The report suggests that the oil shock’s pressure on Chinese costs may be beginning to fade. For global investors, that could ease concerns that imported inflation will force central banks to keep policy restrictive for longer. Still, oil prices and transport conditions will determine whether that improvement holds. In equities, Europe’s rally is attracting increased attention from money managers who see the advance as more than a short-term trade. That is notable after U.S. stocks reached records last week, powered by strong corporate earnings and expectations that weaker jobs data reduces the odds of a September Federal Reserve rate hike. The question for the week ahead is whether investors broaden beyond U.S. mega-cap technology and into overseas markets, industrials, banks, and cyclicals. Technology remains a major driver in Asia as well. Chinese AI-chip designer Moore Threads says it plans a Hong Kong listing at an appropriate time, following a 420% jump in its Shanghai-listed shares since last year’s debut. The potential listing reflects continuing investor appetite for domestic Chinese semiconductor exposure amid trade restrictions and efforts to build alternative supply chains. The broader theme is cautiously constructive risk appetite, supported by earnings and softer inflation signals, but still vulnerable to any setback in Hormuz negotiations or renewed energy-supply disruptions. Thank you for listening to Financial Markets from The Daily FM. See you tomorrow!

Subscribe to this podcast

Latent Space in 3 minutes: The Inference Engineering Masterclass — Philip Kiely & Ali Taha, BasetenHere is The Daily FM summary of the Latent Space that aired on Monday August 3rd. This episode was a deep technical masterclass on inference engineering with Philip Kiely and Ali Taha from Baseten, joining swyx and Vibhu to explain what actually happens after an open model is released and before users experience it as a fast, reliable API. The opening question was simple but revealing: what happens when someone sends a 200,000-token prompt? Philip explained that the system first asks whether part of that input has been seen before. If so, cache-aware routing can send the request to a machine that already has some KV cache, avoiding expensive recomputation. If not, the system may split the work between “prefill” GPUs, which process the giant input and generate the first token, and “decode” GPUs, which produce the output tokens. For coding workloads, Baseten may also use speculative decoding, where a smaller model guesses several tokens ahead and the large model verifies them. That led into the distinction between shared pay-per-token APIs and dedicated deployments. Ali said dedicated deployments become attractive when traffic is high or specialized, because customers can tune batch sizes, quantization levels, routing, and even train a custom speculative decoder for their own traffic. Philip added that dedicated endpoints also avoid noisy-neighbor problems, like someone else benchmarking a shared API with massive traffic. A major theme was that “supporting” a new open model is much more than making it emit one token. Philip said open-source engines like vLLM or SGLang may get basic support quickly, but production readiness requires quantization, calibration, training speculators, testing, routing, and handling new architectural quirks. Ali and Philip gave a striking example from GLM-5.2: Baseten grafted Kimi’s vision encoder onto GLM without changing the language model weights, training only the projector between the “eyes” and the “brain.” Ali said the model learned much better when trained not just to caption images, but to answer detailed questions about them. One of the most surprising sections was on failure modes. Models can collapse into repeating the same token, sometimes not because the weights are bad, but because of inference-engine bugs, CUDA kernel race conditions, or differences between clusters and network interconnects. Ali described cases where the same model behaved differently depending on hardware and KV-cache transfer timing. The discussion on quantization was especially important. Philip framed quality as fidelity to the original full-precision model. Ali explained that quantization is lossy, but Baseten found that quantizing more layers can sometimes preserve quality better, because errors in different layers cancel each other out. They measure this with KL divergence between logit distributions, not just benchmarks, and claimed this can improve throughput by around 20% while maintaining fidelity. The broader takeaway was that inference is still young. Philip said mature fields fight for basis points, while inference optimizations still deliver 20%, 100%, or 200% gains. Stacking NVFP4 quantization, speculative decoding, disaggregated prefill/decode, better kernels, and cache-aware routing can move a model from tens of tokens per second toward several hundred, though the exact gains depend heavily on hardware and traffic. The conversation then widened to NVIDIA Dynamo, model parallelism, mega kernels, Rubin, and AI chips. Philip sees Rubin pushing inference toward systems engineering: moving KV cache around clusters, coordinating GPUs, CPUs, and networks, and designing around memory bandwidth. Ali was notably skeptical of mega kernels, arguing that future GPUs are becoming more specialized and that many fused-kernel approaches may not survive in production. They also covered video generation, where Ali said open-source video still lags far behind closed models like Veo and Kling. The blocker is attention over enormous numbers of video tokens: five seconds can already mean tens of thousands of tokens, and long videos become brutally expensive. Autoregressive video could enable streaming and longer generation, but today’s quality is poor, while diffusion gives better consistency but struggles to scale to long coherent sequences. The episode closed by tying inference back into training. Faster inference helps reinforcement-learning rollouts, while training increasingly has to account for quantization and speculative decoding. Philip predicted continuous loops where deployed models generate traces, get post-trained, A/B tested, and redeployed. Ali gave the memorable example of GLM-5.2 helping profile and write kernels for serving GLM-5.2 itself. The final frontier, they suggested, may be continual learning through persistent KV cache, compacted memory, and models that help optimize the infrastructure they run on. Thank you for listening to Latent Space in 3 minutes from The Daily FM. See you next time!