Personal AI News Digest
37 topics from 3 sources · Archive
3,359 words · ~17 min read
Cursor released a two-year usage report showing median developers generate ~700 lines/week, the 90th percentile reaches ~9,000 lines/week, and the top 1% produce 30-40K lines/week (equivalent to ~45 median developers). Input tokens dominate at a 10:1 ratio (90% input vs 10% output), accounting for ~70% of total AI coding cost despite lower per-token pricing; context caching shifts this dramatically to 90% cache reads, 2.5% cache writes, 7% input tokens, and 0.6% output tokens. Opus 4.7 costs ~10x more per request than Cursor Composer 2.5, but when measured by cost-per-line-accepted, Opus 4.7 matches GPT-5.5 at half the per-request cost while Composer is only 5x as efficient. In one month, the share of Cursor developers allowing AI agents to create commits without manual code review jumped from 10% to ~40%, correlating with Opus 4.7 and GPT-5.5 releases.
Sources The Pragmatic Engineer
Links The Pulse: Interesting AI coding stats from Cursor, Cursor, concluded that writing code by hand is dying
xAI released Grok 4.5, positioned as an Opus-class model optimized for coding and agent workflows. Pricing: $2/1M input tokens, $6/1M output tokens, with cache hits discounted 75% to $0.5/1M tokens; long inputs over 200k tokens cost double. Context window is 500k (down from Grok 4.3's 1M), with Musk indicating a likely upgrade back to 1M within a week. The model represents a significant parameter jump from prior versions and is the first Grok model trained specifically for coding and agents, developed in partnership with Cursor. Day-0 ecosystem support appeared across Grok Build, API, Cursor, Hermes Agent, OpenRouter, and Grok subscriptions, with Cursor offering double usage for the first week. Artificial Analysis ranked Grok 4.5 #4 on its Intelligence Index with a score of 54 (+16 vs Grok 4.3), #4 on GDPval-AA v2 Elo at 1543, and 33% on τ³-Banking (above GPT-5.5's 31%). On the Coding Agent Index it scored 76 in Grok Build, on par with GPT-5.5 in Codex but below Fable 5 in Claude Code. Cost per task: $0.31 for Intelligence Index, $0.49 for GDPval, $2.59 for Coding Agent Index. Average output tokens per Intelligence Index task were ~14k, over 60% lower than Opus 4.8. For Coding Agent Index tasks, Grok 4.5 used 1.9M average total tokens versus 7.2M for Fable 5 in Claude Code and 6.2M for GPT-5.5 in Codex. Artificial Analysis concluded Grok 4.5 sits on the Pareto frontier for cost/performance.
Sources AINews
Links [AINews] SpaceXAI launches Grok 4.5, first Opus-class model post Cursor acquisi…, Grok 4.5, @elonmusk, @elonmusk, @SpaceXAI, @cursor_ai, @cursor_ai, @elonmusk, @scaling01, @ArtificialAnlys
Modal CTO Akshat Bubna discusses the company's $355M Series C and evolution from a developer-focused serverless container platform to an agent-native cloud. Modal was originally built to solve Kubernetes' limitations for bursty, compute-heavy workloads with poor developer experience. The platform has shifted from "Developer Experience" to "Agent Experience," recognizing that agents cannot reason through YAML or dashboards like humans can—they need tight feedback loops, fast iteration, and programmatic infrastructure. Modal's core primitives now include serverless functions, decorator-based infrastructure, elastic inference for custom models, GPU snapshotting, sandboxes, persistent storage, networked containers, private IPv6 networking, RDMA for distributed training, and capacity pools spanning 17 cloud providers. The company works closely with frontier AI labs (Cognition, Ramp, Suno) embedding engineers directly in customer teams. Key insight: agents require observability and hard guardrails more than readable code, and infrastructure must be tighter and more deterministic than what worked for human developers.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO, $355M Series C, Agents need good developer experience too - Modal Blog, Truly Serverless Infra for AI Engineers - with Erik Bernhardsson of Modal
Modal's largest use case is elastic inference for custom models, serving companies like Suno (audio), Runway (video), robotics, and computational biology firms that train proprietary models. These customers face unpredictable traffic patterns—diurnal cycles, product launches, and multi-region deployments with offset scaling cycles. Modal's autoscaling advantage comes from GPU snapshotting, which captures torch.compile model state for faster cold starts, and sophisticated multi-region orchestration. The platform also handles batch jobs (encoding runs requiring thousands of GPUs) and RL rollouts, which can require up to 100,000 sandboxes simultaneously. Unlike frontier LLM inference providers, Modal specializes in the autoscaling problem itself, not just raw GPU capacity.
Sources Latent.Space
Links elastic inference, sandboxes
Modal has open-sourced DeFlash, a block-based speculative decoder that predicts multiple tokens at once rather than one-at-a-time, enabling 2–4× speedup with no quality loss. Speculative decoding works by having a smaller draft model predict tokens ahead, which a larger model then verifies in batch; the key metric is accept length (tokens the big model accepts), and improving accept length gives multiplicative speedups whereas kernel optimizations yield only percentage-point gains. Modal is launching Auto Endpoints, which provide frontier-level inference performance without requiring code changes—users create endpoints via UI or CLI with DeFlash and other optimizations baked in, full code transparency, and the ability to eject into the full Modal SDK for customization. The next step is automatic draft model evolution as data distribution shifts, without manual customer support. Modal's inference advantage over self-hosted vLLM/SGLang setups includes true scaling to zero, elastic burstiness, and production-grade reliability (tail latency control, at-least-once delivery) beyond raw kernel performance.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal built sandboxes in May 2023, before the agent boom, initially demonstrated with smol developer in a self-iterating loop. Sandboxes are isolated, ephemeral compute environments that agents can spin up, inspect, and modify—a natural fit for agent iteration and RL rollouts. The platform now supports networked sandboxes with sidecars (multiple containers per sandbox), outbound networking control (proxies, domain allowlists, credential injection), and multi-node sandboxes. Agents use Modal's CLI for investigation and observability rather than reading code, making observability dashboards and CLI-accessible logs/metrics critical. Modal's sandbox primitives extend to background agents (like Ramp Inspect, which uses snapshotting and fast scaling for reactive behavior) and production agent stacks including persistent storage, networking, and hard security boundaries.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO, sandboxes
Modal supports multi-node training via a decorator that provisions a cluster of GPUs with RDMA (Remote Direct Memory Access) networking at 50 Gbps, bypassing the TCP stack for 3 Tbps internal bandwidth. The platform built private IPv6 overlay networking (I6PN) using eBPF for secure inter-container communication within workspaces, initially for distributed training but now used for other purposes. Modal targets smaller-scale post-training runs (medium-sized foundation models for inference quality) and researcher iteration on large clusters, where elasticity matters more than absolute scale. The company is not pursuing large-scale training runs; instead, it focuses on the elasticity and ease of spinning up multi-node jobs for research and post-training.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal has built internal automation for inference and training research using agents to sweep configurations, run NVIDIA profilers, and optimize hyperparameters and GPU selection (e.g., H200 vs. B200). The company's forward-deployed engineering team (applied inference/training researchers) uses this harness to automate what was previously manual optimization work. Modal Bench is a benchmark suite designed to identify where LLMs struggle with Modal-specific tasks—such as reasoning about observability logs and updating the right configuration—so the team can add skills and surface areas to the product. Auto research is not neural architecture search but guided hyperparameter sweeps informed by model intuition, more efficient than brute-force sweeps. The vision is to make this automation available to customers so agents can autonomously optimize their own inference and training setups.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal has invested in a "compute strategy" team (distinct from FP&A) that handles proactive capacity planning across 17 cloud providers and multiple GPU types. The role involves modeling blend ratios between one-year and three-year reservations, forecasting capacity needs, and making bets on supply chain evolution. The company is launching batch tiers for latency-insensitive workloads (results in 24 hours), targeting computational biology and other batch-heavy use cases that don't need real-time inference. This pricing lever is possible because Modal controls the full stack and scheduling, allowing it to unlock cheaper pricing for customers willing to accept delayed results. The compute strategy role is technically deep, involving financial and supply-chain modeling.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal runs capacity across 17 cloud providers (AWS, GCP, Azure, and numerous neo-cloud/metal providers) rather than building its own data centers, staying capital-light and focusing on software differentiation. The company has invested heavily in a reliability layer on top of heterogeneous providers so that GPU failures or other infrastructure issues don't affect user workloads, enabling Modal to use more capacity than customers could access directly. For real-time audio/video workloads, Modal is building regional routing with fallbacks to place GPUs as close to end users as possible for low latency. Colocation is important for data locality and latency-sensitive workloads (EU, US, Australia pinning).
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal takes a skeptical stance on LLM-mediated permissions at the sandbox level, insisting on hard security boundaries to prevent exfiltration. The company pairs hard guardrails with softer, mediated guardrails. While some argue for trusting a perfect LLM as the kernel, Modal maintains that production agents need explicit, non-negotiable security boundaries. Managed agents (from Anthropic, OpenAI, etc.) are a good starting point for new agent builders, but production-grade agents (like Ramp's accounting agent) need specialized sandbox providers offering fine-grained control over compute, storage, networking, and GPU access.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal serves a diverse set of AI workloads beyond LLM inference, including drug discovery and computational biology (Chai Discoveries), robotics companies deploying models in production, audio (Suno), video (Runway), and real-time multimodal applications. The company's focus on general-purpose primitives (compute, storage, networking) rather than LLM-specific APIs allows it to serve these verticals. Modal is not pursuing an on-premise or air-gapped offering; it remains cloud-only. The diversity of workloads reduces dependence on the LLM inference market and aligns with the company's philosophy of building primitives that work across many use cases.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal deliberately avoids providing simple model APIs (like Replicate's model-as-a-service offering), viewing that market as less sticky and serving primarily hobbyists. Instead, Modal targets companies building differentiated products that need code-level flexibility. Examples include Suno (custom audio model architecture requiring inference tweaks) and companies doing post-training or custom tokenization. The distinction is that Modal's examples are starter code, not black-box APIs—customers can and must modify them. This selection effect means Modal's customers are already building something more differentiated, justifying the deeper infrastructure investment.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal is bullish on the CI/CD market for coding agents, which will run significantly more CI than traditional development. The platform's primitives—memory snapshots and restore—can make CI more efficient by reducing artifact preparation and dependency setup time. The distinction between CI sandboxes (build-time) and runtime sandboxes is that runtime sandboxes have different configuration surfaces (image setup, storage attachment). Modal sees Gitpod/Ona as technically strong but believes they missed the right market at the right time; runtime sandboxes proved more valuable than CI-focused sandboxes as agent use cases took off.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal's runtime is written in Rust and language-agnostic. The company started with Python (the language of data and ML) and has added Go and TypeScript SDKs. Interestingly, agents use the TypeScript SDK more heavily than the Python SDK because agent workloads don't require ML-specific libraries. Modal does not expect to need languages beyond Python and TypeScript in the near term, as these remain dominant for their use cases. The company is not pursuing new language development despite occasional industry calls for LLM-specific languages.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO
Modal has pivoted its SDK team from optimizing developer experience to optimizing agent experience, viewing them as highly aligned (cosine similarity ~0.9). The shift is not just tactical (CLI usability) but strategic: iteration time, spinning up new services, and reducing overhead matter for both developers and agents. Modal Bench identifies where agents lack capabilities (e.g., reasoning about observability logs) and the team adds those as product features or CLI commands. The company believes agent experience is a durable competitive advantage and company-building wedge, as agents will continue to demand faster iteration and lower friction.
Sources Latent.Space
Links Why AI Infrastructure must ev…, Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO, Agents need good developer experience too - Modal Blog
Gergely Orosz answers subscriber questions on AI, engineering, hiring, careers, and the software industry in a live AMA episode. Key topics include: his path from Uber layoffs to founding The Pragmatic Engineer through writing books; his editorial policy shift after learning how Bunq sponsored visa-challenged engineers despite poor working conditions; his investigative piece on Pollen's collapse uncovering unpaid salaries, cancelled health insurance, and a $3.2M customer double-charge with no postmortem; his view that LeetCode-style interviews persist because they select for tolerance of corporate nonsense; his analysis that MCP became an industry standard partly because Anthropic wasn't yet dominant—a feat impossible today given Anthropic's current frontier position; and his stance on AI and skill atrophy: using AI doesn't make work easier if you're pushing hard enough, and he deliberately avoids AI in writing to preserve that skill while accepting hand-coding degradation from AI use. The episode also covers AI's impact on hiring and junior roles, tech debt, engineering manager types, measuring AI productivity, CS degree value, the EU job market, and future-proofing careers.
Sources The Pragmatic Engineer
Links The Pragmatic Engineer AMA, Listen now, The Pragmatic Engineer AMA, YouTube, Spotify, Apple, State of the software engineering job market in 2026, The impact of AI on software engineers in 2026: key trends., How 10 tech companies choose the next generation of dev tools, The reality of tech interviews, chapters
Lilian Weng, cofounder at Thinky, published a comprehensive post summarizing harness engineering literature and its relationship to recursive self-improvement (RSI). The post breaks down proven design trends in harnesses and recaps optimization literature from the ACE paper through recent work like Meta-Harnesses. Key insight: even as harness improvements get internalized into core models, the need to specify goals and context will not disappear. The post frames harness engineering as the central mechanism for self-improvement rather than direct weight modification, a framing that has gained traction across the community including endorsement from figures like Greg Brockman.
Sources AINews
Links Her post, known ACE paper, Meta-Harnesses
Anthropic launched Claude Cowork on mobile and web, repositioning Claude as a background task-running teammate rather than a foreground chat interface. The product emphasizes a shared home tab with tighter Chat/Cowork integration. Separately, Anthropic extended access to Claude Fable 5 on paid plans through July 12.
Sources AINews
Links Claude Cowork coming to mobile and web, @mikeyk, @claudeai
Harness engineering is increasingly the focus of agent architecture design. LangChain launched a Deep Agents course and open-source harness project. Google productized this direction with Gemini API Managed Agents, adding background execution, remote MCP servers, custom function calling, and credential refresh. Sakana's summary connected harness-centric design to The AI Scientist, ShinkaEvolve, and Darwin Gödel Machine.
Sources AINews
Links their thread, @LangChain, @hwchase17, @_philschmid, @OfficialLoganK
Multiple operator-facing agent infrastructure updates: Codex Mobile iOS added task management, filtered diffs, SSH key login, branch comparison, and attachment flows. Hermes Agent added pluggable secrets managers with native 1Password integration and export to private Hugging Face repos. Weaviate 1.38 made its MCP server GA with runtime-gated write access—MCP_SERVER_WRITE_ACCESS_ENABLED can be flipped live without restart. An experimental pattern emerged using Dial MCP server for agents to escalate decisions via phone/SMS/iMessage for human-in-the-loop control.
Sources AINews
Links @Dimillian, @reach_vb, @Teknium’s, threads, @victorialslocum’s post, @omarsar0
Meta Superintelligence Labs launched Muse Image and previewed Muse Video with an explicitly agentic generation loop: planning, web search, tool use, code execution, and self-refinement before rendering. Performance improves with scaled test-time compute, and self-refinement behavior emerged during RL rather than being hand-scripted. Muse Image reached #2 on Image Arena behind GPT Image 2, while Muse Video debuted at #3 on Video Arena.
Sources AINews
Links @alexandr_wang, @_tim_brooks, this follow-up, Arena’s ranking, another Arena post
NVIDIA released Audex, a 30B-parameter model with 3B active parameters in a MoE configuration and 1M context window for unified text and audio work. The core claim is preserving text intelligence while adding broad audio generation and understanding via a single MoE backbone.
Sources AINews
Links @HuggingPapers, @_weiping
Cohere launched Cohere Transcribe Arabic, described as the most accurate open-source Arabic ASR model under Apache 2.0 license, with emphasis on dialects, code-switching, and Arabic-accented English.
Sources AINews
Links @cohere, @JayAlammar
NVIDIA expanded its robotics stack into the Hugging Face ecosystem by bringing GR00T 1.7 and Isaac Teleop into LeRobot for open humanoid robotics workflows. UMA demonstrated a full-stack robotics narrative with a prototype built by a small team in 9 months, emphasizing vertically integrated hardware/software for trustworthy robots.
Sources AINews
Links @NVIDIARobotics’s announcement, integration guide, @RemiCadene, the Northstar reveal, @psermanet’s safety note
Liquid AI released Antidoom, an open-source training method to reduce doom loops where small reasoning models repeat tokens until context exhaustion. The method, FTPO (Final Token Preference Optimization), relabels the loop-triggering token and redistributes probability toward alternatives. Reported reductions: LFM2.5-2.6B from 10.2% → 1.4% and Qwen3.5-4B from 22.9% → 1% under greedy sampling, with downstream eval gains.
Sources AINews
NVIDIA's Puzzle-75B-A9B compression work compresses a hybrid MoE parent model while preserving reasoning, coding, long-context, and agentic quality. Results: roughly 2x server throughput and 1M-context concurrency on H100 rising from 1 request to 8.
Sources AINews
Links @omarsar0
NVIDIA launched Nsight Python 1.0, making GPU performance analysis scriptable in Python.
Sources AINews
Links @HagedornBastian’s post
Unsloth shipped GGUFs for DeepSeek-V4-Flash, plus export to NVFP4/FP8 formats and speedups for GRPO and MoEs.
Sources AINews
Links @danielhanchen’s update
GRPO-style normalization is being adapted for agentic RL at the task or environment level to handle higher reward variance in multi-turn environments. A training-free verifier paper from Stanford/NVIDIA/Berkeley reads calibrated continuous scores off scoring-token logits, posting strong numbers across Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench, suggesting verification is becoming an independent scaling axis.
Sources AINews
Links @cwolferesearch, @omarsar0
Anthropic's J-space work dominated interpretability discussion but drew sharp criticism over consciousness framing. Critics including @danburonline, @paul_cal, and @scaling01 argued the vectors are causal largely by construction under the Jacobian-lens definition. The stronger technical takeaway is cross-model structure: @eliebakouch computed CKA similarity on J-lens geometry across 38 open models and found surprisingly universal layer/depth organization, even across unrelated families like Llama and OLMo. Anthropic and Neuronpedia released J-lens weights for open models. Goodfire introduced Block-Sparse Featurizers for multidimensional concepts in activations, arguing many vision concepts are inherently 2–4 dimensional blocks rather than single directions.
Sources AINews
Links @danburonline, @paul_cal, @scaling01, @jacobandreas, @eliebakouch, this follow-up, their thread
Agent Arena placed Claude Sonnet 5 (Thinking) at #6, with strongest signals in confirmed task success and bash usage, but still with uncertainty around steerability.
Sources AINews
Links Agent Arena
Artificial Analysis launched Harvey LAB-AA, a legal-agent benchmark over 120 private legal tasks across 24 practice areas. Claude Fable 5 led at 14.2% all-pass rate; Claude Opus 4.8 and GLM-5.2 tied at 7.5%, with GLM hitting that at roughly ~6% of Fable's cost per task. The benchmark exposes the gap between passing many individual rubric items and producing acceptable end-to-end deliverables.
Sources AINews
Links their release
Google promoted Experience AI Scientist, a multi-agent system for end-to-end scientific workflows.
Sources AINews
Links this ICML post
DeepMind launched Predicting the Past, grounding Gemini in Aeneas and Ithaca for Greek and Latin historical analysis via plain-English interactions.
Sources AINews
Links their thread
Norm Ai announced a $120M Series C at $1.2B valuation, describing a full-stack "agentic law" setup spanning software plus an AI-native law firm.
Sources AINews
Links @johnjnay’s post
Tencent released the non-preview Hy3 open model collection on Hugging Face, a 295B-parameter MoE with 21B active parameters, now under Apache 2.0 rather than the prior restrictive community license that reportedly excluded use in South Korea, the UK, and the EU. The relicensing removes commercial and geographic usage barriers. Commenters highlighted that claimed benchmark improvements over HY3-Preview may translate to real-world usefulness for high-end local/home inference setups, pending quantized releases and independent testing.
Sources AINews
Links New open model from Tencent Hy: Hy3 (295B total 21B active - apache 2.0), Hugging Face