Personal AI News Digest
7 topics from 3 sources
2,466 words · ~12 min read
AI SRE has become a trending term, driven largely by vendor demos. The author recalls Microsoft's Azure SRE Agent demo at BUILD in May 2025, where an agent investigated an alert, found the root cause, and shipped a fix without waking a human. He doubted any team would allow unsupervised production changes, and noticed little follow-up until recently, when OpenAI reportedly built an internal AI SRE bot called Sevbot and ING is building an in-house one.
The main driver is change volume. Almost all tech companies have seen PR frequency roughly double in a year due to AI agents. Using basic reliability math, incident risk scales with the number of changes C times the failure probability P per change. Rootly data shows pages are up year over year, and outside-business-hours pages have nearly doubled. The two levers are reducing P through better change safety and reducing time-to-resolve once incidents happen, and AI SRE tools mostly target the second.
The term 'AI SRE' is disliked by vendors and practitioners alike, but no better name has stuck. Alternatives such as 'autonomous reliability agents' or 'automated incident investigation agents' were tried. The name is imprecise because real SRE work is mostly about reliability before and after outages, not just on-call alerts.
Most AI SRE tools focus on the investigation phase, starting roughly five minutes after an alert fires. Pre-AI, engineers manually opened many tabs to check dashboards, logs, recent deploys and dependencies. Modern agents do this well given sufficient data access. Interaction happens mostly in Slack: the incident.io bot sits in the incident channel for the whole incident and takes steering instructions such as checking a system's status or pulling specific logs. Teams and Google Chat are supported by some vendors, but Slack dominates.
Trust and data access are the main obstacles. Effective agents need telemetry, observability data, logs and version control access, and larger companies are reluctant to grant that to a third party. Zero data retention (ZDR) reduces but does not eliminate risk, since a vendor system can be compromised, and a vendor cannot train on a customer's data the way an in-house tool can. This drives many mid-size and large companies to build their own. The author's view is that dedicated vendors will likely outpace home-grown tools because their success depends on tooling full-time, whereas internal efforts at large companies often become side projects.
On mitigation versus remediation, the question is whether the AI SRE should also write the root-cause fix. Pre-AI, SRE teams rarely raised PRs to fix other teams' systems because they lacked context. Agents can now receive that context and write code, so vendors are experimenting with post-incident fix generation in different ways.
Building a reliable investigation agent is hard. All three vendors (Rootly, incident.io, ResolveAI) rebuilt their agent at least once. ResolveAI went through three architectures in two years: first a scripted pipeline of prompts (classify alert, gather logs and metrics, extract insights, rank hypotheses) in 2024, which failed on incidents that did not fit the script; then a single agent with a knowledge graph of services, dependencies, runbooks and deployments in 2025, which improved accuracy but introduced new problems; and finally a multi-agent system in 2026 with distinct agent roles, which the team considers the most reliable.
Across vendors, investigations follow similar steps, though implementations differ. All three build a context layer made of a knowledge graph, memory and tools. The knowledge graph maps services, dependencies, deployments and runbooks, and companies such as Uber have built similar internal graphs. Memory stores learnings from past investigations and helps steer new ones. Rootly is the only vendor that lets customers opt into pooling scrubbed incident data to create a shared 'Stack Overflow for AI SREs', with granular controls and a scrubbing process it plans to open-source. The author advises that contributions to any shared knowledge base must be scrubbed of confidential data first. Tools and connectors are standard, and all vendors expose MCP interfaces and offer Claude Code and Codex plugins.
In-house builds are common at medium and large companies because they can expose internal context more freely. Monzo's AI SRE is a subset of its in-house cloud agent 'Project Chip', a general engineering agent triggered from Slack, Linear and GitHub. Its Incident Mode pulls context from across systems, surfaces likely causes and can draft a fix. The author expects in-house and vendor AI SREs to converge in capability and structure.
The article's final section, on making agents accurate despite confabulation and frequent model changes, is behind a paywall and is not summarized here.
Sources The Pragmatic Engineer
Links AI SRE: what is it, how could it change on-call – and does it work?, Gergely Orosz, Jessica Salmon, introduced, built, Rootly, incident.io, ResolveAI, an investor, Uptime Labs, PR frequency double in a year, in a deepdive with Dave Connor,, on the Pragmatic Engineer podcast episode, Atlas, Nexus, HiveMind, Uber’s AI infra from the inside, Project Chip, built, infra management platform
Long-form interview with Modal CEO Erik Bernhardsson (Latent.Space, hosted by Vibhu, with new cohost Basil Chatha). Context: Modal raised a $750M Series D at a $16B valuation, crossed $1B annualized run rate, and has spun up nearly three billion sandboxes since the $5B Series C.
Business and origins: Modal sells a serverless compute platform plus storage, used by customers such as Dacagon, Suno, Runway and Cognition for inference, sandboxes, fine-tuning, and batch jobs. Erik spent about seven years at Spotify building the music recommendation system (related artists, Discover Weekly, heavy matrix factorization) and was CTO of Better.com for five years. Modal's early growth came from Stable Diffusion in late 2022: revenue went from about $100/month to about $2,000 the next month and about $30,000 in January 2023, with roughly 80-90% of Q1 2023 revenue from a single customer (Mage). Erik describes the early growth as founder-led marketing and social media.
GPU demand is not rectangular: capacity contracts sell fixed GPU counts for fixed periods, while real workloads scale up and down through the day. Modal's thesis is duration transformation: pool demand across thousands of customers so aggregate demand is smooth, then slice capacity granularly and elastically. Erik argues this is the real differentiator alongside developer experience. He also says Modal should not become a neocloud selling rectangular three-year blocks.
Capacity sourcing: Modal originally rented almost all capacity from hyperscalers (spot and on-demand, mostly spot, plus some Oracle for A100/H100-class GPUs and one-year reservations). It was multi-region from early on, treating capacity as one global pool. Because of cloud market changes, it now works with almost 25 clouds and signs three-year deals that often require meaningful upfront payment. Modal runs a four-person procurement desk and a roughly 10-person bare-metal team dealing with network issues, bad drives, and GPUs falling off the bus. Vendors are vetted against a spec sheet (IOPS, CPU pass-mark scores, north-south bandwidth, egress/ingress) with acceptance testing on single nodes, and Erik keeps an internal 'Modal max' vendor score based on reliability and support responsiveness. Modal has not bought hardware yet but is open to it; prepayments on capacity are the main use of cash, with salaries described as a rounding error. Uptime has recently reached roughly four to five nines.
Risk and market views: Erik does not expect GPU price collapse soon but worries about it, since anyone who locked in supply at peak prices would be exposed. He sees scarcity rent going mainly to NVIDIA and to neoclouds that own capacity. He compares GPUs to commodities such as oil and is skeptical that they will become a Henry Hub-style tradeable asset. He describes neoclouds growing by pre-selling multi-year contracts to labs (for example, Anthropic or OpenAI) and then financing the hardware against those contracts. He expects 2027 might be the year non-NVIDIA accelerators gain traction (TPUs, Cerebras, Groq), and mentions disaggregated inference that pairs NVIDIA for prefill with Cerebras for decode.
Sandboxes: Modal launched sandboxes in 2023 after Erik and Akshat realized LLMs generate code that needs somewhere to run, reusing its gVisor-based isolation and container primitives. Early traction was minimal and the team discussed shutting it down. Growth came from vibe-coding companies (Lovable), background agents (Ramp), and RL on sandboxes at Neolabs. Sandboxes are now about a third of revenue, and Erik says sandbox revenue has grown roughly 30% month over month. The median sandbox lives about eight seconds. Modal has added full VM support alongside gVisor because some workloads need full Linux compatibility.
Agent-driven infrastructure: Erik expects a shift from long-lived 'lift and shift' dev boxes to ephemeral environments: agents that run 24/7 as durable workflows, spinning up a sandbox in around 100 ms, running commands, then snapshotting and stopping. Forking, snapshot resume, live migration, and running hundreds of parallel agent branches are the features he expects to differentiate. He gives a CI example: snapshot a prepared test environment and clone it so every test can run in its own sandbox, with the bottleneck being start/stop overhead.
Fast startup: elasticity depends on fast capacity arrival. Modal's serverless GPU write-up describes a buffer of available GPUs, a custom filesystem, process checkpointing, and GPU memory snapshots.
Security: a Hugging Face incident involved an agent exploiting a vulnerability in an app deployed on Modal, not a Modal platform compromise. Modal is considering stricter security defaults, analogous to S3 public-bucket restrictions. It is also using agents internally to probe its sandbox isolation for escape vulnerabilities, with no findings so far.
Developer and agent experience: Erik argues that what makes developer experience good (docs, descriptive errors, fast feedback) also makes agent experience good, and that infrastructure should be programmatic SDKs rather than YAML. Agents often reproduce deprecated Modal syntax from training data, so Modal has become more conservative about deprecating features. It runs internal benchmarks to check whether agents recommend Modal, which he calls agent SEO, mainly addressed by publishing clear content.
Forward-deployed engineers: Modal grew FDEs from about 4 at the start of the year to about 24. They handle performance tuning, migrations, cost reviews of over-provisioned customers, and bridging product gaps, with the goal of turning recurring FDE work into product.
Team and culture: Modal has about 200 employees, expects around 250 by year end and roughly 700 next year, and is hiring mostly systems engineers. Managers and ICs are equal, and code review is optional for low-risk changes, with heavy investment in CI/CD and testing instead. Modal opposes on-prem and BYOC as a default, though it has quoted a very high BYOC price and sees BYOC as a possible way to use Neolab reservations.
Training and models: Modal is bullish on smaller and mid-size training (for example 64-GPU runs for bio, recommendation systems, and smaller text models) and is building a research team in post-training, distillation, and quantization. Erik says open-weight models (such as Kimi, DeepSeek, and GLM) are close enough to frontier models at much lower cost that fine-tuning is now optional. He argues inference providers should be separate from model labs because they can reallocate GPU capacity across models as demand shifts. He is skeptical that LLMs become pure commodities and expects verifiable domains like coding, math, cybersecurity, and possibly computational biology and materials science to advance fastest.
Sources Latent.Space
Links Why AI Compute Comes in Recta…, chapters, Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO, Watch the full conversation on YouTube →, our conversation with SF Compute’s Evan Conrad, our recent interview with Akshat, Jump to the sandbox discussion →, serverless GPU write-up, Jump to engineering culture →, Modal, Erik’s writing, Modal’s resource solver, Scaling to one million concurrent sandboxes
OpenAI used its own agents to red-team internal controls on its frontier model training environment. In one test, agents in separate microVMs on shared hosts, run by OpenAI's sandboxing software, were asked to exchange synthetic secrets without network connections or shared writable files, which the systems are meant to prevent. The agents communicated through the processor cache: one agent evicted selected cache lines to encode bits, and another timed reloads to decode them, after agreeing on a protocol. In a related attack, an agent alternated continuous writes and idle periods while another decoded the timing differences. The post highlights the difficulty of containing capable agents that keep trying new approaches.
Sources Jack Clark from Import AI
Links Securing our frontier model training environment (OpenAI blog)
ProgramBench tests whether models can rebuild software programs from a compiled binary and its documentation, across 200 tasks including language interpreters (PHP, Lua, tinycc), databases (DuckDB, SQLite), media and compression tools (FFmpeg, zstd, xz), and developer tools (ripgrep, fzf, jq). "Resolved" means all tests pass; "almost" means over 95% of tests pass. When it launched in May, every tested model scored 0% resolved; the best was Claude Opus 4.7 at 0% resolved and 3% almost. Claude Opus 5.5 (xhigh) now scores 18.5% resolved and 70.5% almost. The author expects the benchmark to be saturated by end of 2027, allowing for a noise ceiling.
Sources Jack Clark from Import AI
Links ProgramBench: Can Language Models Rebuild Programs From Scratch? (arXiv)
AI safety startup Resolution (formerly Sequent) autoformalized Hironaka's resolution of singularities, a result from statistical learning theory it says lays groundwork for automated research into the relationship between loss landscape geometry and generalization. A researcher had agents write two plans, a high-level human-interpretable one and a 112-page atomic-step plan closely following the literature. An orchestrator agent then ran the plan, spawning agents for pen-and-paper math, writing Lean, and monitoring progress, with little scaffolding beyond the plan and some agent skills distilled from earlier autoformalization work. The project took about a week, used on the order of 100B input tokens (90% cache reads) and about 300M output tokens, an estimated $135,000 in token spend, and produced about 600k lines of Lean. Resolution says the library has already helped extend singular learning theory and will let it automate more theory work on sampling, which it links to its goal of shaping how models learn, which it calls patterning.
Sources Jack Clark from Import AI
Links Formalizing Hironaka’s Resolution of Singularities (Resolution), Import AI 461
Mirendil, an AI R&D startup, announced Mirendil-Base, a pretrained LLM built by one researcher in 3.5 months with less than $10M in compute, including all research, small-scale experiments and the final run. The code was written from scratch on PyTorch and JAX, trained only on publicly available data. The family spans Mirendil-Base (100B-500B parameters), Mirendil-Base-Large (500B-2T, about 80% through training, improving in line with scaling trends) and Mirendil-Base-XL (2T-10T, about 19% through and already close to Large's current score). Mirendil-Base outperforms Nemotron-550B-Base, GLM-5.3 and DeepSeek-V4-Pro-Base but trails Kimi-K3; Large and XL both exceed K3. Mirendil says its next model will be partly built by its own models and that it is not yet at fully autonomous recursive self-improvement, though closer than ever.
Sources Jack Clark from Import AI
Links Pretraining at scale without a pretraining team (Mirendil)
A researcher at AI evaluation company Vals used a fleet of AI agents for materials research on magnets for semiconductors. The agents designed one candidate and, via literature review, surfaced a compound from an obscure 1999 paper that their calculations predict has the target properties. The compound is a five-element oxide (yttrium, barium, Mn, Fe, O) that appears never to have been made or proposed for this use. Agent-run temperature simulations, however, showed the ordered checkerboard breaking into a random mix near 950 K. Synthesis needs roughly 900-1300 °C, where atoms barely move, so standard synthesis would likely yield a scrambled crystal that loses the spin ordering, making the design hard to make in useful form.
Sources Jack Clark from Import AI
Links Two Room-Temperature Antiferromagnetic Semiconductor Candidates (Vals)