Personal AI News Digest
17 topics from 2 sources
3,705 words · ~19 min read
The essay argues that as models absorb more of the agent harness into their weights, the remaining harness faces the human and becomes an "Attention Interface": the layer that decides when the agent interrupts you, what it shows you, and which decisions it needs you to make. Claude Mods are an early, user-programmable version of this layer. A mod is a Claude Code plugin made of TypeScript or JavaScript functions that run inside Claude Code in the CLI and Desktop. Mods respond to events such as prompts, tool calls, and interface renders. They can observe an event, change how it is rendered, or handle it in place of the default behavior, so they work as runtime middleware that changes Claude Code's execution and rendering flow. The author's bet is that the harness becomes a platform for interfaces between people and agents, similar to how the browser became a platform for web apps.
Claude Code has a built-in skill for building mods, so a user can describe the mod in plain language and Claude builds it. Anthropic ships six mods as built-ins, visible under /plugin > Installed > Built-in. Most run in the background. The author highlights two. You Should Know is an Anthropic side-agent, off by default, that watches the main agent and flags consequential actions buried in the trace. In the author's case it caught that CompareView.swift and about 45 other files already had uncommitted work, which would have been lost if the demo edits were discarded. Next Steps Supervisor, part of the open-source ModSmith plugin, runs after turns that edit files, run commands, or make 8 or more tool calls. It surfaces flagged items and offers the suggested next step as a clickable button that sends the instruction back to Claude Code.
The essay compares this with two other products. OpenAI's Intelligent UI, shipped with GPT-6 in ChatGPT on October 7, lets the model generate charts, maps, checklists, forms, and small tools on the fly from a component library OpenAI built; the main user control at launch was turning visuals down. Cursor's PR Routing & Approval assigns reviewers by code ownership and commit history, auto-approves low-risk PRs against a risk-score threshold, and combines Bugbot and Security reviewer findings with repo policy files. If a PR edits an approval policy, the policy is read from the base branch, so an agent cannot relax its own review. Cursor's docs say the feature "does not replace a full code review." The author's distinction is that ChatGPT and Cursor give users an interface someone else designed, while Claude Mods let users design their own with Claude, at the cost of having to make it useful.
The essay lists three downsides. Noise: Next Steps Supervisor ran on 99 of 111 turns after its October 15 install, so users skim past it. Cost: each check forks and re-reads the conversation, roughly 5-6k input-equivalent tokens per check on a 50k-token session per the mod's README, more when the cache has expired. Trust: mods run with the user's permissions, can read and write files, see every prompt, and spend usage; Anthropic's docs say mods are not sandboxed, and built-in mods keep running under --safe-mode. The author recommends reading any third-party mod's code before installing it. Mods are easy to fix, though: /supervisor off disables it for a session.
The author's test for a mod is whether it surfaced something he would have missed, made a decision easier, or let him act without distraction. Today the human still notices the friction and asks for the mod. The proposed next step is a background process that reads session transcripts, finds repeated corrections or manual checks, suggests a mod, and helps test and refine it from feedback. The essay closes by asking readers which friction with Claude Code a mod could solve today.
Sources Latent.Space
Links Claude Mods: Flipping the Harness Inside Out, Claude Mods, The Evolution of the Agent Harness, built-in skill, six of them built-in, turn it on, ModSmith, Intelligent UI, PR Routing & Approval, Anthropic’s docs, Claude Code’s docs on creating mods
DeepSeek open-sourced DeepInfer, the inference engine behind V4.1 Flash, a customization of vLLM that credits SGLang and TensorRT-LLM. It supports distributed inference, speculative decoding, SWA Bounded Replay and structured output on NVIDIA SM100 and Huawei Ascend 950. Released with it is deepseek-recipe, a Rust library with Python bindings for Chat Completions, Responses and Messages protocol conversion and parsing. Not released: the KV-cache management and load-balancing components, which remain coupled to internal infrastructure, plus prefill/decode disaggregation. Disclosed production figures (first-party, unaudited): 155T tokens/day per cluster, 96% input-cache hit, 1.5M cache-inclusive input tokens/s per prefill GPU, about 14,000 output tokens/s per decode GPU over roughly 2,400 GPUs, 400 ms average TTFT, 230 tokens/s generation (300 for coding). Prefill routing does not need cache affinity because cached prefixes live in a shared 3FS cluster (about 100 PiB SSD, 3-day retention for agent traffic, 7 days for chatbot traffic). Decode routing balances request count and KV length. Expert-parallel load balancing replicates popular experts. SWA Bounded Replay reconstructs short-lived state from the latest 128 tokens and reportedly nearly halves prefill compute. DSpark adaptive speculative verification gives a 6% throughput gain over fixed five-token verification. vLLM's response: decode throughput within 6% of DeepInfer's published results on GB200 under a matched configuration, plus AgentX results and reproduction recipes, though this is decode parity, not reproduction of the full service.
Sources AINews
Links [AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/c…, DeepInfer, DeepSeek releases DeepInfer, New from DeepSeek - DeepInfer -, GitHub, Deepseek released DeepInfer - Open Source high-performance, large-scale inferen…, deepseek-recipe, zhuanlan.zhihu.com/p/2095560295930508953, customization of vLLM with credit to SGLang and TRRTLLM, 3FS, published it here, research guide, explainers, codemap, subsystem guide, @vllm_project, @vllm_project, reproduction resources
OpenAI added 12-hour API prompt caching, which can be mixed within one request with the 30-minute default. Longer retention carries extra cost but benefits repeatedly reused agent context. OpenAI also disclosed that about 40,000 accounts had applied usage limits that did not match their selected plans, corrected the issue and promised credits.
Sources AINews
Links 12 hours, announcement, announcement, configuration, incident update, @OpenAIDevs, retention controls
Perplexity's Local Computer lets the model and agent run entirely on supported Apple-silicon Macs, including offline, using PPLX Pro (a post-trained Qwen 3.8 version) with no Computer credits, for Pro, Max and Enterprise subscribers; 36 GB unified memory is recommended. A hybrid mode splits work between local and cloud models, so its privacy depends on the split. LangChain's Managed Deep Agents package per-user credentials, scoped memory, sandboxed execution and scheduled work. Hugging Face's Agent Manager orchestrates persistent sessions across Claude Code, Codex and other harnesses.
Sources AINews
Links announcement, model, requirements, LangChain architecture, Agent Manager
SGLang merged megakernel support; SemiAnalysis reports nearly 1,000 tokens/s/user on GB300 NVL72 in AgentX and 1.5x greater interactivity on MI355X. Inferact reports 391 tokens/s/user for Kimi K3 on TPU v7 using open-source megakernels in vLLM TorchTPU. Zyphra reports GLM-5.3 512K-context TTFT falling from 178.8 s to 64.6 s on MI355X, or 42.3 s to 11.8 s with 90 percent cached context; these are single-node tests without live traffic. PyTorch 2.15 replaces several MPSGraph paths with Metal kernels: 5 to 6.5x faster MoE decoding, Qwen3.6-35B-A3B rising from 9.4 to 61.3 tokens/s on M3 Ultra, FP8 on MPS, and Wan2.2 VAE peak driver memory falling from about 270 GiB to 8 GiB.
Sources AINews
Links SGLang result, TPU result, results, conditions, release summary
Tenex, an applied AI partner founded about 18 months ago and now 75 people (roughly 50% Fortune 2000 and 50% mid-market firms with $100M to $500M revenue), pays engineers by story-point output rather than hours. Alex Lieberman says the feared failure modes (sandbagging estimates, shipping bad code to rack up points) have not materialized, because AI strategists are paid on customer retention and growth and so push back on work that does not serve the client. The harder problem is compensating non-ticket work such as architecture, client discovery and deciding what to build. CJ Hess says his time spent directly coding has dropped as models improved, and more time now goes into getting customer context into a form the model can use. Pay spread across the firm is about 4x between highest and lowest paid engineers, not 10x.
Output per engineer has roughly doubled this year by Alex's estimate, with the inflection dated to around December and Opus 4.5 (long-running task capability). Tenex calibrates story-point value about once a quarter, and Alex says the 2x increase has not been fully reflected in point values. Opus 5.5 is the team's default driver, with Codex used for review and back-and-forth on the local machine. Tenex uses AMP, and says hooking in a Claude Code subscription cut costs by roughly 80%. Token spend relative to employee spend is about 1:1 to 2:1. Costs have fallen since the team stopped defaulting to the most expensive frontier models, and CJ argues that a per-task cost decline has made runaway token spend much less of a risk than six to eight months ago.
MetaHarness is the core of their workflow: a shared set of skills, a canonical plan document, an opinionated decision/activity log, and a CLI that agents use to drive planning, visualizations and client readouts. Everything is checked into Git so that the reasoning behind decisions persists in the repo rather than in people's heads. Enforcement of the document structure has been reduced as models improved. Tenex's own article notes this structure can be overkill for a small team on one project.
On enterprise adoption, Alex's rough numbers across about 100 enterprises: about 90% have rolled out Claude or Codex company-wide, about 60% have done training and enablement, about 50% have coding agents rolled out to engineering and product, and only about 5% have agents in production for non-engineering work. The most common ask is back-office agents, and the recurring blocker is permissioning and RBAC for agents, which Tenex now productizes as its own gateway. Alex's point is that you cannot automate a process that does not exist, so standardizing workflows and building evals comes first.
On vertical AI, Alex is generally bearish, arguing that one-of-one expectations push software toward services margins and that frontier models may absorb tasks. CJ's counterpoint is that domain data, specialized interfaces and workflow integration (Harvey, OpenEvidence, Bloomberg as the analogy) create durable moats. CJ distinguishes a vertical product whose main feature is an agent doing a task from a tool that improves the workflow around that task. The discussion also covered whether the interface will consolidate into a few agents that users already have, with SaaS vendors offering APIs and MCPs rather than their own agents.
In Kill My SaaS, the host replaced a conference speaker-management tool quoted at roughly $40,000 to $43,000 a year for about $1,000 of build cost. Alex's view is that ripping out software gets harder further upmarket, and that for mid-tier tools the question is whether the saving justifies the ongoing maintenance burden. The host's argument is that owning the workflow allows iteration in minutes rather than waiting months on a vendor. Alex also argues that AI-driven automation has to be paired with reallocation of staff, and that employees have agency in reskilling.
The interview closes with hiring for forward-deployed engineers and AI strategists, and a brief discussion of AI culture in New York versus San Francisco.
Sources Latent.Space
Links Latent Space: The AI Engineer…, AI for the Fortune 2000: How …, chapters, Listen now, AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman &…, Alex Lieberman and his cofounder joined us, Kill My SaaS, Jump to the compensation discussion →, AI-native SDLC article, CJ explains the workflow →, Watch the exchange →, Watch the debate →, Tenex, Alex Lieberman on Tenex, Latent.Space
Latent Space interview with Evan Conrad, CEO of SF Compute, recorded at AIE NYC. The conversation covers how a GPU compute market is built, how physical delivery works, and what must be true before the market becomes liquid.
Origin and physical settlement: SF Compute started by trying to sublease a cluster. Conrad describes the hotel analogy: each customer request (cleaning, room service, security) pulled the company further into operating the whole building, until it was effectively running a cloud. The company has since built much of the operating capacity of a cloud so it can deliver usable compute, not just a cash payment against an index. The term 'physical settlement' is borrowed from futures markets, where the buyer receives the underlying (oil, power). The oil-went-negative-in-2020 episode is cited as an example of what happens when delivery logistics are ignored. Conrad argues the order book and physical settlement had to be co-designed, so that real constraints such as interruptibility, storage recovery time and networking show up in the contract.
Market maturity and liquidity: Only about 5% of total capacity is on the order book, but liquidity there is concentrated. A customer who spends $100 buying capacity makes back about $25 on average by reselling it, and this is one of the metrics the company tracks to judge whether liquidity solves the customer's problem.
Heterogeneity and standardization: Standardized contracts are still far off. Conrad gives an example of a single operator offering two VM types with a 40% performance difference, caused by GPU kernel launch latency that degrades MFU. Instead of standardizing first, SF Compute built a matching engine where orders carry many attributes (GPU generation, memory, NVLink domain size, time window, interruptibility, storage) and are matched against sell orders with those properties. This generates liquidity early, at the cost of wider bid-ask spreads, which creates arbitrage for buyers who can use many capacity types. Customers care about GPU generation because of VRAM and NVLink domain size. Cohort effects also matter: early H100 and GB200 batches had firmware and hardware faults (some traced to vendors such as Supermicro), so buyers may pay more for later-deployed units. Conrad says the company is trying to reduce dimensionality while keeping SKUs open to be defined by others, and says the 'named query' feature is new.
Term structure and pricing: Contracts range from an hour out to three to five years. Clusters are typically financed over three to five years. For blocks, customers currently pay fully upfront, while longer leases are handled differently. Conrad distinguishes two kinds of market: forward capacity that is not yet online, and capacity currently in use, where selling it means interrupting the running workload. Batch inference can tolerate interruption with about an hour's notice, while live inference with SLAs needs weeks or months of notice, so the two are priced differently. Conrad says the forward curve shows a hump: spot prices fall as blocks approach delivery and cannot be sold in the past, then rise further out. He cites a large arbitrage: a B300 can earn roughly $40/hour running inference at full utilization (about $25/hour at 50%), while B300s can be bought for about $5-6/hour.
Independence and incentives: Conrad argues an exchange operator must not own hardware, or it could manipulate the forward market, so SF Compute must be 'Switzerland'. He argues that NVIDIA benefits from the market existing but cannot run it neutrally, and that hyperscalers, neoclouds and inference providers have demand-side conflicts that stop them from doing it. He also warns against financialization before physical settlement exists, citing the credit default swap market (described as about 12 times the size of the underlying bonds) and LIBOR manipulation as cautionary examples. He says indexes are often built from thin data with weak correlation to the actual price of buying new reserved capacity.
Ecosystem and operations: SF Compute works with Silicon Data (cash settlement, which he thinks is valuable if derived from physical settlement) and mentions other players doing standards (SemiAnalysis), residual value insurance, and lending. The stack covers site visits, bill of materials, supply chain with OEMs, cluster architecture, rack rollout, and backplane management (PDUs, switches, DPUs, UFM, BMC). The company does not self-perform builds and works with partners. It targets sovereign compute projects worldwide.
Milestones: Conrad reports several hundred million dollars in closed volume and says the company can now subdivide InfiniBand fabric on the fly without contacting the provider, which helps platform-as-a-service customers scale clusters up and down. He describes this as virtualizing the data center and making it programmable.
Sources Latent.Space
Links The Programmable Data Center …, chapters, Listen now, The Programmable Data Center — Evan Conrad, SF Compute, the new Compute Trading track at AIE NYC, former guest Martin Casado, software-defined networking, commoditizing compute, SF Compute, Evan’s AI Engineer NYC talk: How Compute Markets Work, Why oil futures went negative in 2020 — EIA, How SF Compute customers resell unused GPU capacity, SF Compute’s $245M compute sale, Abel: SF Compute’s settlement ledger, CFTC: Barclays and LIBOR manipulation, Elastic InfiniBand partitions
A user reports Claude Opus 5.5 completed Baldur's Gate 3 autonomously via a harness of claude -p with built-in tools disabled, a Python MCP server connected to a Lua mod, Vulkan-layer screenshots and SDL-mediated clicks. The run took 39 hours and 712 USD across 99 fresh sessions, 20,809 tool calls, 862 screenshots and over 2.2B cached input tokens; a git-backed memory system held 157 commits. An audited rerun took 36 minutes and 24 USD. The author plans a short-scenario BG3 benchmark scored by success, attempts, time and cost. Separately, a Kotaku report claims Opus 5.5 helped produce browser-playable ports of PS2-era games such as Halo CE and GTA Vice City, but commenters argue these are recompilations rather than true decompilations.
Sources AINews
Links After $712 and 39 hours, Opus 5.5 beat Baldur’s Gate 3 all by itself, write-up, code, Claude Opus 5.5 decompiles a PS2-era game overnight, and hundreds of them now r…, Kotaku report
DeepSeek disclosed about 38 USD revenue per accelerator-hour at off-peak prices. Revenue is not profit, and the figure excludes SSD cache fleet, CPUs, networking, idle capacity, failures and staff. An initial community estimate of 87.76 to 96.94 percent margin assumed 97 percent cached input, 2 percent uncached input and 1 percent output of billed tokens, 80K uncached input tokens/s and 16K output tokens/s per GPU, and 3 to 6 USD per GPU-hour rental. That estimate was revised: the idealized case implied about 49 USD/hour versus DeepSeek's 38 USD/hour, and the revised margin range was 84.2 to 96.1 percent. At 7 USD/hour against 38 USD revenue, the arithmetic gives about 81.6 percent before other costs. At 5 to 10 USD/hour the accelerator-only margin spans about 73.7 to 86.8 percent. Yuchen Jin notes DeepSeek did not disclose the GPU behind the figure. Extrapolations (about 166B USD per GW, 81-day payback, roughly 100B USD annual revenue from about 4,167 racks) are speculative scenarios, not disclosures. Some see profitability; others expect token-price declines to compress returns.
Sources AINews
Links point out in their additional disclosure, off peak prices, >80% inference margin, initial estimate, revised estimate, industry extrapolation, @KCDN19, price-convergence argument, hosting competition
Arena ranks V4.1 Flash 20261017 Max as the top open model in Agent Arena and Code Arena WebDev, at #11 overall in Agent Arena. It shows +5.1% net improvement across 7,500+ real-world sessions at a 0.20 USD median cost per task, which Arena describes as about 37 percent of GPT-6.1 Sol's improvement at 27 percent of its cost. Other signals: +11.5% confirmed success, +5.9% praise-versus-complaint, +5.4% Bash reliability, +2.5% steerability, no tool hallucinations observed. Vals reports V4.1 Flash 1017 exceeded GPT-6 Sol's average progress on a simulated corporate-network task using a Pi-based agent, with three runs each for DeepSeek and Sol and single runs for others. DeepSeek used more tokens but had lower total API cost. Critics note DeepSeek used about 4.3x the tokens and 2.4x the time, and that Sol's best run reached an objective no DeepSeek run reached.
Sources AINews
Links @arena, detailed results, @ValsAI, cost interpretation, methodology, @scaling01, @teortaxesTex, equal-cost swarm suggestion
Artificial Analysis launched leaderboards for 5,000 general health conversations and 525 clinician conversations. GPT-6 Astra leads Professional's length-adjusted score at 66.0 percent; Claude Sonnet 5.5 leads general HealthBench at 62.1 percent. Opus 5.5 has the highest raw Professional score at 71.4 percent but drops to 59.9 percent after length adjustment. DeepSeek V4.1 Flash scores 60.9 percent at 0.005 USD per task versus Sonnet's 62.1 percent at 0.08 USD per task. Grading uses OpenAI models.
Sources AINews
Muse Glimmer is an open 30B multimodal agent model distilled from Muse Spark, targeting 32 GB VRAM and running on one RTX 5090 with 4-bit quantization and a speculative decoding head. The paper reports 75.5 on MCPAtlas and 51.2 on SWE-Bench Pro. Qwen3.6-27B remains ahead on OSWorld-Verified and TerminalBench 2.1. It supports text, images and more than 100 languages under Apache 2.0.
Sources AINews
Links paper summary
Mistral launched Voxtral 3 with batch and streaming transcription, multilingual TTS and a voice-agent SDK with tool calling, available via API or on customer infrastructure. Artificial Analysis places Small Transcribe 3 at #3 with 2.1% AA-WER. The realtime model reaches 3.3% final-transcript AA-WER at 960 ms. TTS enters the top 12 in both voice arenas, priced at 13 USD per million characters, generating 112.3 characters/s in measurements. Kyutai separately released Pocket TTS variants down to 1.9M parameters, claiming 103x realtime on one CPU core with voice cloning.
Sources AINews
Links launch, deployment options, evaluation, throughput
Bria's Fibo 2.0 is an 8B open-weight image family with Dev for text-to-image and distilled Edit Turbo. Artificial Analysis reports leading open-weight editing performance and a statistical tie with Qwen-Image-2.1 for text-to-image. Bria API pricing is 32 USD per 1,000 images. The weights are non-commercial; commercial use requires a separate Bria agreement. Krea 3 Ultra ranks #5 per Artificial Analysis (earlier posts said #4), and Arena ranks it #4 in multi-image editing and #8 in single-image editing, so the defensible claim is competitive editing quality.
Sources AINews
Links Diffusers, fal, Artificial Analysis, Arena
Anthropic and Citrini report 17 binders from 211 tested protein designs, after substantial computational work by an internal Claude research model, and are releasing all tested designs and lab data. None bound multiple toxins and none were tested for neutralization, so this is not an antivenom. Sophont's Holon-S1 predicts spatial gene activity and biomarkers from routine pathology slides, evaluated across 16 organ types and more than 15 scanners, outperforming two comparison models; these are research findings, not validated treatment-selection results. In an evaluation, GPT-6 Astra and Claude Fable 5.1 were each given unpublished AI-research questions and a 10,000 USD budget. Original authors rejected both generated papers for weaknesses in creativity, judgment, calibration and communication, though experimental design and resource management improved.
Sources AINews
Links announcement, data, researcher account, official caveat, launch, evaluation
OpenAI announced the acquisition of Orchestra Bio, a drug-development coordination platform, extending its life-sciences work into operational execution alongside Rosalind; terms were not disclosed. Bob McGrew's Arda launched manufacturing software developed with Anduril as a design partner; the founder account specifies a 160M USD Series A at a 1.5B USD valuation, while CNBC cites 230M USD raised without reconciling the figure. Reports say OpenAI is seeking about 30B USD at a 1.4T+ USD valuation ahead of an anticipated Anthropic IPO. A reported WSJ review of Anthropic's draft prospectus cites 16.3B USD first-half revenue, nearly 9B USD compute spending and more than 600B USD in future compute commitments. These are attributed reports, not completed transactions.
Sources AINews
Links announcement, founder announcement, CNBC report, OpenAI report, Anthropic report
METR agrees with Anthropic's low catastrophic-misalignment-risk conclusion for covered models but identifies significant analytical disagreements. It recommends end-to-end red-teaming of monitoring systems and assessing whether models could undermine safety-relevant decisions. Hugging Face CEO Clement Delangue criticized Netflix's portrayal of a reported OpenAI-agent cyberattack, arguing containment, monitoring, defensive use of open models and missing agent traces deserved more attention.
Sources AINews
Links review, recommendations, decision risk, statement