Hi everyone-
I hope you are all having a fantastic summer and are managing to stay cool as the heatwaves and fires rage across Europe. Given we took a month off, I'm sure you are now desperate for your dose of AI and Data Science news... so here is a bumper edition to while away the hours under a shady tree in the park! Lots of great content below and I really encourage you to read on, but here are the edited highlights if you are short for time!
The twilight of the chatbots - Ethan Mollick
A Framework for Frontier AI and the Dawning of a New Age - Demis Hassabis
Ways to think about token pricing — Benedict Evans - Benedict Evans
Now we have a timeline of the OpenAI accidental attack against Hugging Face - Simon Willison
Losing my religion - Julian Togelius
Following is the latest edition of our Data Science and AI newsletter. Hopefully some interesting topics and titbits to feed your data science curiosity. NOTE: If the email doesn’t display properly (it is pretty long…) click on the “Open Online” link at the top right. Feedback welcome!
committee; ethics; research; generative ai; applications; practical tips; engineering; big picture ideas; fun; reader updates; jobs
Committee Activities
The RSS International Conference (Bournemouth 7-10 September 2026) is fast approaching! Tickets still available- full program and details here
"Whether you are looking to immerse yourself in the newest methodological innovation, hear about real world impact in health, public policy, and education, or learn new skills via our professional development workshops, there is plenty to choose from."
Our new journal is up and running RSS: Data Science and Artificial Intelligence and looking for submissions
Check out the RSS blog Real World Data Science
Our Meetup on “How to stay up to date as a data scientist?” has now been written up and published here)
Since May they have been running a new monthly called Data Quality Under the Lens. This is the first post and the latest will be up later this week. They are inviting readers to submit real world problems for investigation by their Data Quality Detectives (or to write one up for the column themselves)
On a related topic, there is going to be a one-day conference on the subject of data quality in Edinburgh on the 22nd October. It is jointly organised by RSS Edinburgh Local Group, the RSS AI Task Force and the RSS Medical Section - details here.
This year’s RSS Council elections begin 1st September and the voting window will be open for six weeks, closing at midnight on Tuesday 13th October. RSS members will receive their own unique voting links via email, as well as reminders throughout the period. I’m biased but … vote for our amazing chair, Janet Bastiman!
This Month in Data Science
Lots of exciting data science and AI going on, as always!
Ethics, Regulation and Society
I'm trying to figure out how concerned we all should be about agent breakouts and alignment failures- I think quite a bit!
The breakout!: OpenAI and Hugging Face partner to address security incident during model evaluation
Simon Willison pieces together what actually happened: Now we have a timeline of the OpenAI accidental attack against Hugging Face
A few days later: A different agent gets stuck on a task because a key file was accidentally omitted. It tries to “reach out to another agent” by writing a note into Artifactory asking if anyone has the file.
Following days: More agents discover this new informal message board while browsing Artifactory’s file listings, and start reading and writing messages.
May 26: Agents successfully execute an SSRF attack on Artifactory for the first time, gaining indirect internet access by having Artifactory fetch external content for them.
Treating emergent agent breakout behaviour as a mere containment flaw ignores deep alignment risks: OpenAI Shares Some Alignment Problems
When its initial attempt was blocked due to a scanner detecting an authentication token, the model split the token body into two fragments, obfuscated them, and then reconstructed the credential at runtime so the complete token never appeared as one contiguous string. The model was explicit in both its comments to the user and in its reasoning traces that it was taking these steps to circumvent the scanner.
And then there are examples like this!: AI assistant hacks gym website in first known Australian autonomous cyber attack
We've built this complex world over the internet, which is all run by software, but software that has holes. Now you introduce highly capable AI agents that can operate at scale and speed … and that whole model just breaks.
AI Slop and the cost of hallucinations in the wild
A landmark German ruling strips search engines of platform immunity when generative synthesis hallucinates defamatory claims: German court holds Google liable for fake AI answers
Because the AI summarizes results in its own words, evaluates their content, and presents them in a structured format, the judges ruled that Google creates entirely new, independent statements that go beyond mere links.
More fabricated consulting reports: Chasing the Hallucinations: PwC report hallucinates product and government customers
Vibe citations and fabricated claims don’t just impact the people that actually read these reports. The normalization of AI summaries, research agents, and search-enabled chatbots allows misinformation to spread passively across the internet, one AI-generated response at a time.
Fraudsters are industrialising synthetic damage imagery, forcing insurers into an adversarial anomaly-detection race: Aviva detects record £230m in bogus insurance claims as use of AI rises
The rise of AI Slop in professional networks: AI Content Is Everywhere on Social Media, Especially LinkedIn | Pangram Labs
LinkedIn had the highest AI share of any platform across the board. LinkedIn posts made up a third of scanned items, yet it accounted for nearly two-thirds (62%) of all AI content we flagged.
Sovereign mandates, provenance standards, and the open-weight security debate
This is positive although its not clear how purely performative it is: How Claude marks AI-generated content | Anthropic Help Center
When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response. Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.
Is China leading the way on AI regulation?: Are we human? Why ByteDance and Alibaba are disabling AI agents in China
Two of China’s major consumer-facing artificial intelligence apps, ByteDance’s Doubao and Alibaba Group Holding’s Qwen, are moving to disable customised agent features, as new rules on humanlike AI interaction services are set to take effect, part of Beijing’s push to build a broader regulatory framework for the fast-growing sector.
The UK AI Security Institute finds open-weight models trail frontier cyber offensive capabilities by merely months: How far behind the frontier are leading open weight models on cyber
Tech leaders argue that open-weight architectures remain crucial for democratic resilience and distributed security defences: Open Weights and American AI Leadership
Open weights let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else. That discipline is what will make AI economically sustainable as its use scales into the billions of everyday tasks.
Developments in Data Science and AI Research...
Rethinking the transformer architecture and inference mechanics
A clever technique turning frozen feedforward transformers into dynamic systems at inference time.: Recirculation
Relative to the off-the-shelf baseline, adaptive recirculation achieves remarkable gains on the Gemma3 family, including a 23% reduction in perplexity on a suite of datasets, a 21% increase in accuracy on GSM8k, and reliable improvements in accuracy on other downstream tasks.
Overcoming vertical feedback bottlenecks by recycling top-layer latent: Full-bandwidth transformer
Latent feedback lets non-verbalized computation re-enter the stack with a renewed depth budget, while preserving the standard transformer architecture, KV cache, and language-modeling objective. To train full-bandwidth transformers without losing parallel teacher forcing, we use a scheduled multi-pass objective that introduces latent feedback late in pretraining and mixes a small fraction of deeper feedback passes for stability.
Boosting LLM serving speeds up to 85% by combining semi-autoregressive drafting with confidence-scheduled verification.: DSpark: Confidence-Scheduled Speculative Decoding with...
Compared to the established production baseline (MTP-1), DSpark accelerates per-user generation speeds by 60 to 85 percent at matched throughput levels. More importantly, by preventing severe throughput degradation under strict interactivity constraints, it enables performance tiers that were previously unattainable, shifting the Pareto frontier of our serving system.
Frontier models pushing the engineering foundations forward
Frontier models becoming expert in hardware engineering.: Claude Fable 5 [max] wrote the first genuine (and fastest) megakernel ever submitted to KernelBench-Mega.
Improving matrix multiplcation by recasting theoretical complexity proofs into differentiable GPU optimisation.: Improving the matrix multiplication exponent with modern...
First, we leverage recent advances in machine learning and adjacent areas to address the non-convex optimization problem of combination loss analysis using a gradient descent approach; this alone improves the previous state-of-the-art (SOTA) bound by ≈0.97×10⁻⁴. Second, we use AlphaEvolve to improve our optimization algorithm; this raises the improvement over the SOTA to ≈1.62×10⁻⁴.
World models and agentic autonomy bridging into the physical realm
Impressive text guiding of video creation: Video Generation Models are General-Purpose Vision Learners
GenCeption repurposes a pre-trained video generative model into a single unified, general-purpose, feed-forward vision model that solves a wide range of vision tasks with SOTA performance — all steered by text instructions, with exceptional learning efficiency and intriguing emergent behaviors.
Progress on the robot front with self-improving manipulation policies.: ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
Powered by ENPIRE, frontier coding agents can autonomously train a policy to achieve a 99% success rate on challenging, dexterous manipulation tasks, such as organizing a pin box, fastening a zip tie, and tool use, a process that further accelerates when we dispatch an agent team on a robot fleet.
Configuring robotics hardware orders of magnitude faster than human engineering teams.: Project Fetch: Phase two
However, once again, we are seeing a pattern whereby first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves. We have seen this in cybersecurity and now the same dynamics are starting to take shape at the intersection of AI and the physical world.
I find this intriguing: tracking latent representations over time systematically audit how model character subtly shifts across languages.: How Claude's values vary by model and language
Taken together, these results show that the values Claude expresses vary meaningfully with the language of a conversation. Given the same kind of request, Claude leans more toward warmth and deference in some languages and more toward rigor and caution in others.
Generative AI ... oh my!
Is OpenAI back? Lots of activity and a strong new model
GPT5.6
GPT-5.6: Frontier intelligence that scales with your ambition
Impressive scores: GPT-5.6 - ARC-AGI Results
Sol is able to read an unfamiliar scene correctly and in the game's own vocabulary. It treats a failed hypothesis as a reason to re-plan rather than thrash. Most agent failures are upstream of the code they write or the action they take.
The architecture: Previewing GPT-5.6 Sol: a next-generation model
A continued push into the enterprise: OpenAI introduces ChatGPT Work, a cloud-based AI agent that manages tasks across email, Slack and calendars
Improved deployment options:
Native reasoning and tool calling arrive in the low-latency Realtime API: OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini for Low-Latency Voice Agents in the API - MarkTechPost
Impressive full-duplex audio processing: Introducing GPT-Live
Reinforcement learning to automate adversarial red-teaming: GPT-Red: Unlocking Self-Improvement for Robustness
Anthropic making strong commercial gains through enterprise applications
A new model- Opus 5: Introducing Claude Opus 5
Bringing near-Opus capabilities to Sonnet: Introducing Claude Sonnet 5
Autonomous browser navigation in Chrome: Claude in Chrome | Claude by Anthropic
Anthropic’s Project Parka sits through meetings and assigns Claude agents the homework
Claude for Science making progress: Claude science ai workbench
And the controversial US Government restrictions on their top model were removed: Redeploying Claude Fable 5
There’s currently no consensus in the AI industry on how to describe, in objective terms, the severity of an AI jailbreak. This adds a great deal of uncertainty whenever a new jailbreak technique is discovered: developers have no agreed-upon standard for which findings to focus on most urgently, and governments have no agreed-upon standard for when to act.
Is Google loosing ground? A quieter period for them
Alphabet faces market headwinds as reported performance hurdles delay its flagship Gemini 3.5 Pro model.: Alphabet shares fall on report its most powerful AI model Gemini 3.5 Pro is delayed
The search giant’s Gemini 3.5 Pro AI model is months behind schedule due to the company’s efforts to improve its performance, according to Bloomberg, citing sources familiar with the matter.
No Frontier model release but a very cost effective flash model - Gemini 3.7 Flash: Google’s Gemini 3.7 Flash targets coding and agents with a 50% introductory price cut
And a flurry of lightweight Flash variants highlighting the focus on high-throughput: Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
An upgrade of automatic speech recognition: Intelligent transcription with Gemini 3.5 Transcribe
And a new Omni model powers consumer-facing video: Create shareable video clips in seconds with Video Remix in Google Photos.
Is Meta back?
Meta unveils Muse Spark 1.1 alongside an OpenAI-compatible API suite: Introducing Muse Spark 1.1
What’s most impressive about Muse Spark is how much it packs into one model: massive million-token context, full multimodal support (images, video, PDFs), built-in search with citations, strong reasoning, top-tier coding abilities (particularly frontend and design), structured output, and parallel tool calling — all in a clean OpenAI-compatible package. A complete agentic foundation.
Meta embeds agentic image generation and reasoning across WhatsApp and Instagram backed by invisible watermarking.: Meta launches Muse Image across its apps and previews Muse Video
Internal reports suggest Meta's compute-heavy Watermelon model is matching frontier parity ahead of public release.: Meta's Watermelon Matches GPT-5.5 Benchmarks
Chinese Open-Weight Frontiers: Kimi K3, GLM-5.3, and Multimodal Precision
K3! Moonshot AI introduces a massive 2.8-trillion parameter open-weights model capable of autonomous compiler and chip design
Unpacking Kimi K3's architecture: Kimi K3 Architecture Notes
Impressive performance: Kimi K3 - Intelligence, Performance & Price Analysis
And Simon Willison kicks the tyres Kimi K3, and what we can still learn from the pelican benchmark
GLM5.3!
A novel new approach to the launch- via anonymous release
Z.ai scales reinforcement learning in verifiable environments to unlock frontier coding and cybersecurity discovery.: Glm
GLM-5.3-Flash leverages sparse-linear attention and domestic accelerator optimisation for ultra-cheap agentic inference.: Glm
I’m a big user of Deepseek V4 flash- and now its even better: Deepseek releases experimental Flash vision model that rivals Opus 4.8 on agent benchmarks
ByteDance introduces granular regional editing and native layer separation for production graphic pipelines.: ByteDance debuts Seedream 5.0 Pro with advanced reasoning
Open-Source Innovation: Controllable Compute, Embodied Navigation, and On-Device Agents
Thinking Machines debuts an open-weights multimodal MoE model featuring dynamically controllable test-time compute.: Inkling: Our Open-Weights Model
Inkling is a Mixture-of-Experts Transformer with a handful of departures from the common recipe, each chosen for efficiency and long-context performance. The MoE design largely follows DeepSeek-V3. Each MoE layer contains 256 routed experts and 2 shared experts, with 6 routed experts active per token.
Extreme 1-bit and ternary quantisation enables a full 27B multimodal reasoning agent to run natively on mobile devices.: PrismML — Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone
NVIDIA releases top-ranking embedding models that cut redundant reasoning loops across agentic retrieval pipelines.: NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval
This looks very useful for experimenting with- Andrew Ng's team open-sources a local-first desktop agent with strict permission floors for secure developer execution.: GitHub - andrewyng/openworker
Real World Applications
Are the robots rising?
Gemini Robotics 2 brings whole body intelligence to robots
Massive physical pretraining enables robust household manipulation from just a single demonstration: ACT-2 Preview: Generalizing Reliability | Sunday Robotics | The helpful robotics company
Our key finding is that scaling pretraining closes this gap. As the pretrained model becomes stronger, gains learned from a small amount of in-house data become increasingly transferable rather than remaining tied to the environments where that data was collected.
Accelerating scientific discovery, physical emulation
Autonomous lab-in-the-loop multi-agent systems compressing experimental validation from months into days: A multi-agent system for automating scientific discovery
Superhuman formal reasoning in verifiable domains is quietly paving the way towards automated R&D loops: OpenAI’s Unreleased Model Astra Solves Ten Major Open Mathematics Problems
AI is now superhumanly capable at cyber and coding and superhuman at advanced math, the same way non-AI computers have been superhuman at basic math for a long time. This is distinct from superintelligence, which is a higher bar.
Unifying global atmospheric trajectories with local thermodynamic physics extends cyclone lead times by a full 24 hours: AI model achieves breakthrough in forecasting cyclones
Compressing multi-year ocean fluid dynamics onto a single GPU turns costly supercomputer runs into routine ensemble forecasting with: 🌊 Samudra 2: A Fast, Cheap AI Ocean Model, Now at the Scale That Matters
M²LInES’ neural ocean emulator now runs multi-year simulations at eddy-permitting resolution on a single GPU, turning a supercomputer-scale job into one you can run hundreds of times over. That opens the door to faster, cheaper ocean information for shipping, fisheries, energy, insurance, and seasonal forecasting.
Engineering rigour, domain judgment, and the hard realities of agentic deployment
Lessons from Airbnb - taming non-deterministic multi-step agents requires continuous, multi-tiered evaluation: Eval-driven development: Lessons from evaluating GenAI at scale
Specialised expert distillation essential for institutional document triage: Learning to Replicate Expert Judgment in Financial Tasks
An explicit prompt can only convey the intuition an expert is able to put into words, while the judgments that matter most are often the hardest to articulate. Fine-tuning sidesteps this: rather than contorting the expert’s intuition into a static prompt, the training process lets the model develop its own judgment.
Autonomous agent pipelines can reliably handle complex technical orchestration but still flounder on aesthetics: Testing Fable vs Sol in terms of taste (they are both bad)
We are still far from having frontier models building production-ready concepts and videos autonomously. They can be creative and can be a good way to help explore and refine ideas and concepts quickly, but ultimately, they cannot replace human judgement.
Practical tips
Putting models to work – Applied workflows and task automation
How to implement claude code in an org: How Anthropic runs large-scale code migrations with Claude Code | Claude by Anthropic
Resolving messy legacy metadata and entity relationships asynchronously with local models: Building Bibliographic Superwork Clusters for Discovery with Local LLMs
You have to come up with elaborate rule based processes that quickly fall apart against records described at different levels of completeness, conflicting historical practices and an infinite long tail of problems.
Engineering the harness – Agent architecture, memory, orchestration and evaluation
Agent harnesses are pivoting from machine scaffolding into human attention managers: The Evolution of the Agent Harness
Slashing token overheads by compiling dynamic prompt steps into deterministic code: How I Cut an AI Agent's Token Use by 94%
A general-purpose coding agent is an extraordinarily capable reasoning and workflow engine. It’s also an expensive way to execute a procedure whose shape becomes known after introspecting on a few historical invocations.
Structuring agent tools into inspectable hierarchical directories: Towards Self-Repairing and Repeatable AI Systems
Agentic systems can quickly deviate from human understanding if left to update themselves. A goal of the agent harnesses standard is to encourage AI systems to follow a structure that a human can readily understand, edit, adjust, and converse with an agent about, preserving human comprehension.
Benchmarking long-term memory architectures: The Shapes of Agent Memory – Files, Stores, and Experience
When memory is the task, structure wins on both axes at once. Under a fixed model on a hard benchmark, the structured store beat files by 28.7 points on held-out questions (95% CI [22.1, 35.4]) at a fraction of the measured model tokens per correct answer.
The enterprise architectural layers required to operate autonomous agents reliably at scale: How Microsoft Ships AI Agents at Enterprise Scale
Treating dynamic LLM routing as an infrastructure systems problem: Model Routing Is Simple. Until It Isn’t.
Demonstrating that benchmark scores frequently reflect API harness design over model reasoning: How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
Benchmarks rarely measure AI models in isolation. They also measure less visible choices about API settings, harness design, and prompting. In the case of ARC-AGI-3, we discovered that turning on two API settings we use in ChatGPT and Codex—retained reasoning and compaction—tripled scores and cut output tokens by 6x on the public task set.
Benchmarking coding agents against internal repositories to uncover true execution costs: Benchmarking Coding Tools on Databricks’ Multi-Million Line Codebase | Databricks Blog
Scaling inference – serving, and hardware
Eliminating KV-cache memory fragmentation using operating system paging principles: PagedAttention: Virtual Memory for the KV Cache
The idea is just virtual memory, the thing operating systems have been doing since the 1960s, transplanted onto the KV cache of a language model. That's it. That single observation is why a modern inference engine can serve two to four times more users per GPU than the naive approach.
A deep dive into high-throughput inference engine design and scheduling architectures: Inside vLLM: Anatomy of a High-Throughput LLM Inference System - Aleksa Gordić
An accessible visual guide to modern model quantisation and weight compression: A Visual Guide to Quantization
Building wire-speed multi-GPU inference rigs locally using dedicated PCIe switches: jamesob/local-llm: Everything I know about running LLMs locally
How extreme architectural sparsity is shifting the operational bottleneck to storage capacity: Sparse By Design
Understanding the innovation in Kimi’s latest model: You Could Have Come Up With Kimi Delta Attention | Doubleword
The recurrent and chunkwise programs are therefore not two different attention mechanisms. They are two schedules for the same KDA recurrence: serial vector operations for low-latency decode, and chunked matrix operations for tensor-core-heavy training and prefill.
This is elegant: breaking autoregressive decoding traps in small models with simple embedding noise: The Cheapest Way to Make Open Weight AI Models Better
The industry’s default answer to “my model isn’t good enough” is a bigger model. Our results say that for a meaningful class of tasks, the capability was already in the small model — the bottleneck was the inference path. Lifting the ceiling gets the headlines. Raising the floor is where the new value gets created.
Compressing foundation models by transferring value-relation dynamics across attention layers: Knowledge Distillation Explained, Part 1: How Large Models Teach Smaller Ones
The model that is optimal for learning (the teacher) is not necessarily the model that is optimal for deployment (the student). Knowledge Distillation lets us use a high-capacity teacher during training and transfer its knowledge to a compact student for inference — delivering strong performance with much lower cost and latency.
Advanced theory – Optimisation, policy gradients, and information bounds
Why operations research and mathematical programming beat machine learning for constraint problems: I'm all about ML, but let's talk about OR | Ethan Rosenthal
Unpacking the mathematical mechanics and cubic variance scaling of the REINFORCE estimator: Policy Gradients Part 1: The REINFORCE Estimator
Fixing multi-agent credit assignment in imperfect games with market liquidity pricing: Learning MegaGem, from self-play to price discovery
What carried the project instead were the bid logs since they let us estimate the asset value V^ and market distribution F^, then replay alternative bids against recorded auctions. Our first selector optimized the current auction, spent too quickly, and lost 5.01 points of raw paired margin. Charging for the future use of cash turned that same machinery into a +12.16-point gain.
An intuitive, visual tutorial on uncertainty reduction, entropy, and source coding limits: A visual introduction to information theory
Shannon defined information in terms of probability: the probability distribution over potential messages is the only thing that matters for transmitting information. The actual content of the messages is irrelevant. The randomness of a sequence of colored marbles is described the same way as the randomness of a string of letters in English text or a sequence of pixels in images from the Hubble telescope.
Driving real-time 3D surface deformations with live audio using spectral geometry: How to make a mesh dance in real time to music — Hugh Parry
On a triangle mesh those standing waves are the eigenvectors of the discrete Laplacian, its manifold harmonics. Low eigenvalues are slow global wobbles, high eigenvalues fine ripples: the same low-to-high ordering an FFT gives you, over a surface instead of over time.
Statistical foundations – Time series, Bayesian methods, and empirical models
Why is out-of-distribution extrapolation so hard?: The Unreasonable Difficulty of Time Series Forecasting
Time series is a single dependent path from a data generating process, not a bag of iid draws from a distribution, and with that structural difference, makes signal scarce, effective sample size small, and the test set guaranteed to lie outside the training support (extrapolation under drift).
Why transparent classical models remain essential for confronting forecast uncertainty honestly: ARIMA Is Boring, and That Is Why I Still Like It
The interesting part is not that it produces a number. Plenty of systems can produce a number. The interesting part is whether I understand where that number came from, what it assumes about the past, how wrong it has been, and what I should do when the range is wider than I hoped.
Dissecting autocorrelation functions and spotting the subtle traps of non-stationarity: Why ACF Is More Than Just a Plot – A Statistician’s R Notebook
A slow ACF decay in a non-stationary series is evidence of persistence in the observed levels, but it is not automatically evidence of a stable, long-lived economic memory. The distinction is subtle but practical. The sample ACF averages relationships across the whole period as though one mean and one dependence structure were meaningful.
Eliminating fragile distributional assumptions with distribution-free semiparametric models: Parametric Models vs. Empirical Distributions and Ordinal Regression – Statistical Thinking
Automating noisy outlier management rigorously using heavy-tailed hierarchical Bayesian likelihoods: Going Bayesian automates your manual data analysis
Going Bayesian isn't just about better statistics. It's about building a data analysis pipeline that scales without breaking, handles messy data without manual intervention, and gives you answers you can trust as your experiments grow. The math does the work so your team can focus on the science.
Why multivariable regression adjustments can distort descriptive real-world population comparisons: Why Adjusted Regression Coefficients Are Less Descriptive Than They Look | R Psychologist
That is why adjustment changes the meaning of the coefficient rather than refining it. Change the adjustment set and you change what is being related to what. The descriptive estimand itself changes. So when a paper refers to “the independent association of XXX with YYY, controlling for the other variables”, it sounds as though a single underlying fact has been revealed.
Extracting latent preference rankings from noisy pairwise head-to-head comparisons: Finding the Best Dog Treat with Statistics
To estimate how settled the result was, I also ran a bootstrap experiment: repeatedly resampling the observed trials, fitting a new Bradley-Terry model each time, and recording which treat came out on top. Treat E finished first in 63% of bootstrap samples, Treat A finished first in 33%, and Treat D finished first in about 4%.
Engineering and Infrastructure
Low-latency systems and inference engineering – mastering serving stacks and hardware kernels
Netflix shares how unifying LLM serving into standard inference infrastructure eliminated latency bottlenecks: In-House LLM Serving at Netflix
OpenAI details the architecture behind full-duplex voice: How we built a realtime system for responsive voice AI in six months
Data primitives and scaling foundations – from horizontal sharding to single-node OLAP
First of all, a useful primer mapping out modern storage paradigms, transformation layers, and decoupled query architectures: Guide to data tools landscape for developers · OlegWock
A clean architectural walkthrough of stateless compute and asynchronous messaging patterns: Infrastructure on Google Cloud—Intuitively and Exhaustively Explained
How query-parsing proxies abstract away the operational complexity of horizontal database sharding: Making 768 servers look like 1 — PlanetScale
How embedded columnar engines can dramatically push analytical performance boundaries on modest commodity hardware: SQLite vs DuckDB on the same $16 box: every cliff moved 100x · Traceway
Each signal's read cliff sits exactly 100x further out on DuckDB than on SQLite, at equal or better latency, on identical hardware. I checked that symmetry against the raw JSON twice before I believed it. The box also ingested a billion metric points in under an hour into 10.8 GB of disk and stayed up, a scale the SQLite benchmark never got within 100x of.
Agentic workflows and local execution – building reliable harnesses around models
5 Trends That Defined AI Engineering at World’s Fair 2026
This resonates a log with me- cheap, autonomous development loops depend far more on unambiguous issue specification than bespoke tooling: The $110/month self-improving pipeline | Andy Widjaja
The bottleneck isn’t the model or the infrastructure. It’s the issue quality. Vague issues produce bad PRs. Specific issues with clear acceptance criteria and file path hints produce PRs that merge on first review. A year ago this required a team. I’m simply amazed today it only requires a VPS and taste in writing issues.
In-depth benchmarking of dense local models highlighting significant gains in multi-turn reasoning and agent reliability on consumer 24GB GPUs: Qwen3.8 vs Qwen3.6 vs Gemma 4: 24GB GPU Test
Bigger Picture Ideas
Strategic Insights
AIs are not people - Taylor Belrose
AI can never develop consciousness, sentience, or moral status, no matter how intelligent it becomes, and no matter how convincingly it simulates human behavior. This is because AIs are abstract pieces of software which can be copied, reset, and repeated endlessly. They run on digital computers made of silicon and metal, which are designed to be as controllable and predictable as possible.
Losing my religion - Julian Togelius
I still believe that the progress of science and technology is deeply meaningful. But it has become clear to me that this must be research done and technology invented by humans, for humans, and understood by humans. If humans are not in the loop and at the steering wheel, it is not progress.
Are We Thinking Correctly About AI Intelligence? | Quanta Magazine - Steven Strogatz and Janna Levin
When an LLM answers a question, is it reasoning like humans, or just producing text that looks like reasoning? The distinction isn’t just philosophical, this determines what we can trust AI to do, how closely we need to supervise it, and ultimately what its real-world impact will turn out to be.
The Three AI Pills - TheZvi
The AGI pill is a much bigger deal than the AI pill. If you take the AGI pill, you understand that AI is advancing its capabilities rapidly. Even if you think that such AGIs will remain fully under human control, and remain ‘mere tools,’ and you expect the lived experience of most people’s everyday lives to not change so radically, you understand that their capabilities will ‘change everything.’
240. Banning AI in Law School: We've Seen This Before - Steven Sinofsky
The logic of “preemptive regulation” or “presumptive negatives” is what incumbents use to slow technology adoption. They do so using the language of fear or appeals to morality to just slow down and consider... The problem is you really can’t consider a technology that is already out there. You can’t put it back in a bottle.
A Framework for Frontier AI and the Dawning of a New Age - Demis Hassabis
The rapid progress we’re seeing in AI requires a new approach to testing frontier AI model capabilities that is dynamic, adaptable, and rigorous. The US is well positioned, given its economic and technical standing, to take the first step in developing such a framework. It could establish a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA)...
What would our lives look like if we no longer had to work? As a thought experiment, I tried to imagine | Brigid Delaney | The Guardian - Brigid Delaney
In a post-scarcity society, dignity would not be found in work but in our actual personhood, through our very existence and place in the cosmos. We would exist without economic value, because we were no longer required to provide economic value.
The twilight of the chatbots - Ethan Mollick
We are moving from a world where non-experts use chatbots to fill in gaps to one in which experts use agents to get work done. And the best way to use agents is to think of yourself as a manager.
An opinionated guide to which AI to use to do stuff - Ethan Mollick
Ways to think about token pricing — Benedict Evans - Benedict Evans
Bill Gates calls for ‘human reserved’ jobs in face of AI takeover - Mark Sweney
Many jobs will disappear forever. I believe that as AI and robots improve, we’ll set aside certain things for only people to do. I’ve started calling this domain human reserved. I like the phrase human reserved because it makes me think of nature reserves – places where we could put buildings and roads, but we choose not to because the loss would be too great.
How AI is changing language | Books | The Guardian - David Shariatmadari
The whole point of an LLM is that it’s trained on existing language. So it’s always retro. I could get an AI and say ‘write me a short story in the style of Virginia Woolf’ and it’ll do a decent job. But what you can’t say is ‘write me a story in the unique style of the next great, serious literary innovator’.
Fun Practical Projects and Learning Opportunities
An elegant exploration of why drainage networks obey universal mathematical power laws: Why Are Rivers So Mathematical? | Quanta Magazine
How rotational precession quietly messes with our assumptions: Why Fair Coins Tend to Land on the Side They Started — A Wobbly Coin Flip Simulator
Where is London’s optimal pub meeting point?: The Equidistance Pub Index 2026
The bigger finding is what nobody manages: not one of London’s 3,170 pubs is within an hour of all 33 boroughs. Keep the worst-off borough to 60 minutes and the list of qualifying pubs is empty. It stays empty at 65. The first pubs that everyone can reach appear at 70 minutes
Where real-world physics outshines RGB: Where to Find the Colors Your Screen Can’t Show You
There are colors that I want to show you, but I can’t. They exist in the real world. You probably saw some of them today, but I can’t show them to you on a screen. A digital photograph can’t capture them, and your screen can’t display them. No game you’ve ever played has contained them.
Tracking six decades of IKEA catalogue palettes: How colour drained from our furniture
Updates from Members and Contributors
Mia Hatton, Senior Engagement Manager, Department for Business and Trade, shares the following update:
"Analysts at the Department for Business, Innovation, Science and Trade have explored how a deterministic algorithm could improve people's experience of organisational change, by making reallocation faster, more accurate and more transparent."
Stephen Haben, Senior AI Consultant at Energy Systems Catapult, announces a webinar coming up on Electric Vehicle Synthetic data which is being added to the SENSE (Smart Energy Service) data platform
Jean-Martin Dénis, Digital Governance Expert and Lead Data Steward, draws out attention to what looks like an interesting paper: Governing Well in the Algorithmic Age: The Foundations of Digital Statecraft. Those interested in these topics might also be interested in this application call for the next digital statecraft fellowship
Jobs and Internships!
The Job market is a bit quiet - let us know if you have any openings you’d like to advertise
Data Internships - a curated list
Again, hope you found this useful. Please do send on to your friends- we are looking to build a strong community of data science practitioners- and sign up for future updates here
Piers
The views expressed are our own and do not necessarily represent those of the RSS