If Turing Post is part of your weekly routine, please share it with one smart friend. It’s the simplest way to keep the Monday digests free. |
|
| To celebrate crossing 100,000 readers, we’re offering 20% off the annual Premium subscription. The offer ends March 31. | | You will immediately get access to our most popular: | | | We are also seeing growing interest in consulting sessions with us. At the moment, you can choose whom you would like to speak with: Ksenia Se or Will Schenk. Get a consultation here. |
|
| This Week in Turing Post: | | | Main topic: Why “Robots Will Take Our Jobs” Is Lazy Thinking | Last week I went to school conferences – it is when your child presents what he or she was up to this quarter or semester. And my son, who is 7 years old, started declaring to me how dangerous it is that robots can steal human jobs. I was smiling politely, but inside I was raging: how is it possible that we feed our children such one-sided information? We have to raise people who can ask questions and see different angles. So I am writing this editorial to send to school. Maybe they will show it to the kids. At least, to discuss. | So, what is wrong with the statement that robots will steal human jobs? | The first thing wrong with it is that it sneaks in bad economics and presents it as common sense. Economists even have a name for this mistake: the lump-of-labor fallacy (Marc Andreessen posted a long thread about it). It is the idea that there is a fixed amount of work in the economy, so if a machine does more, a human must do less. But economies do not work like a pie with a predetermined number of slices. Technology can remove some tasks, lower costs, create new demand, and open entirely new kinds of work. That does not mean every worker benefits automatically, and it certainly does not mean transitions are painless, but it does mean the simple story children are being told is wrong at the level of first principles. | Take ATMs. In the United States, the number of ATMs quadrupled from roughly 100,000 to 400,000 between 1990 and 2010. If the classroom slogan were true, bank tellers should have disappeared. They did not. Teller employment actually rose modestly from about 500,000 to 550,000 between 1980 and 2010. Why? Because ATMs made branches cheaper to run, branches expanded, and tellers shifted away from routine cash handling toward customer-facing sales and service. Same sector, different task mix. James Bessen’s research on computerization reaches a similarly inconvenient conclusion: occupations that use computers have tended to grow faster, not slower, even though the transition often requires new skills and can widen inequality for workers who are not prepared for the shift. | But there is an even bigger point that the classroom slogan misses entirely. Before computers, there were no software engineers, UX designers, cloud architects, or cybersecurity analysts. Before the internet economy, there were no creators, app developers, search marketers, or platform moderators. Before renewable energy became an industry, there were no solar panel installers or wind turbine technicians. The work did not exist because the surrounding system did not exist yet. So with AI – and for that matter robots – we simply do not know yet what new professions will appear. How about we talk with kids about this more? | Of course, the opposite fairy tale is wrong too. Automation can absolutely destroy specific jobs and damage specific communities. The Luddites were reacting to real disruption, and later waves of mechanization displaced farm labor, switchboard operators, and plenty of other occupations. Research by Daron Acemoglu and Pascual Restrepo makes the point clearly: automation has a displacement effect, but history also shows a reinstatement effect, when technology creates new tasks in which human labor still has an advantage. In one striking example, when AT&T mechanized switching between 1920 and 1940, operator jobs fell, but future cohorts did not end up with lower overall employment because demand rose in clerical and service work. The serious question, then, is never whether one machine replaces one person. The serious question is whether society is creating new tasks, new skills, and new pathways fast enough for people to move into them. | And this is where the current moment becomes almost absurd. We are warning children about a future with no work while living in an economy that still struggles to find enough workers. In January 2026, the United States had 6.9 million job openings, and there were about 1.06 unemployed people per opening. Health care and social assistance alone had 1.41 million openings. In February 2026, labor-force participation was 62.0%. At the same time, the Congressional Budget Office projects slower population growth and an older population ahead. | That means: a big part of the problem over the next decades is not too many workers chasing too few jobs. It is too few workers, in the wrong places, with the wrong training, trying to enter a labor market whose task mix is changing. | So no, I do not want my children taught that robots will steal jobs and humans will lose paychecks. I want them taught something harder and much more useful. Technology changes the shape of work. It can enrich societies and it can punish workers. It can raise productivity while leaving some people behind. Whether people keep getting paychecks depends less on a robot existing than on whether schools, businesses, and governments help people learn, adapt, bargain, and move into the new work that technology creates. Children should leave school with curiosity, range, and enough intellectual independence to ask better questions than the adults in the room. Teaching fear is the lazy version of education; teaching children how work actually changes would be far closer to the real thing. | I mean, you don’t have to like AI, but you at least need to understand the bigger picture and how the economy works. | | Topic 2: Beyond LLMs: JEPA and the Road to AGI – the main milestones so far (Is this a story about “who wins”? Watch the episode!) |  | JEPA: The Other Path to AGI / Let’s Map those JEPAs out! |
|
| | | |
| | We are reading/watching/learning: | Agentic AI and the next intelligence explosion by James Evans, Benjamin Bratton, and Blaise Agüera y Arcas. This research from Google argues intelligence doesn’t scale as a monolith but as a society: networks of humans and AI agents debating, collaborating, and self-organizing. Cheng Lou found a load-bearing wall that everyone assumed was structural, and showed it could be removed. That kind of unlock doesn't come along often in front-end! →
| | News from the usual suspects ™ | Z.ai – Clawing Back Control
Z.ai unveils AutoClaw, a local-first agent runner promising full autonomy: no API keys, no cloud dependency, and no data leaving your machine. With support for custom models and its own GLM-5-Turbo tuned for tool use, it’s a clear pitch to privacy-conscious developers. Early chatter praises the freedom – but some raise eyebrows over uninstall persistence quirks. Stripe – API Keys, Without the Archaeology
Patrick Collison introduced Stripe Projects, a developer preview aimed at sparing agents and humans alike from the ceremonial misery of account creation, dashboard spelunking, and API-key scavenger hunts. From the CLI, developers can provision services like PostHog in one command, with billing and credentials handled in the background. A neat idea: less “modern stack assembly,” more actual building. OpenClaw – Maturing even more
OpenClaw’s 2026.3.28 release is a small municipal reform. New provider support, approval hooks, richer channel bindings, image tooling, and a formidable parade of fixes suggest a project intent on being everywhere at once – and, annoyingly for competitors, working there too. OpenReward – Reinforcement Learning, Now With a Proper Marketplace
OpenReward is making a bid to become the clearinghouse for RL environments: 330-plus benchmarks, hosted infrastructure, simple APIs, and enough task variety to keep any frontier lab pleasantly overcommitted. The pitch is elegant – open standard underneath, managed convenience on top. In other words, freedom for the principled, and speed for the impatient. A rare startup instinct: have both.
| | 🔦 Models and Agents Highlight | Speech generation and voice models | Voxtral TTS – Delivers a strong multilingual text-to-speech and voice-cloning result with very short reference audio, making it relevant for anyone tracking the convergence of expressive speech generation, compact prompting, and open-weight voice systems. →read the paper Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale – Expands the “generalist plus specialist” idea into science by combining strong general multimodal ability with deep task coverage across scientific domains, making it notable both as a scaling story and as a domain-foundation-model play. →read the paper UI-Voyager: A Self-Evolving GUI Agent Learning via Failed Experience – Learns from failed mobile interaction trajectories rather than pretending only successes matter, which is interesting because GUI automation remains one of the clearest real-world tests of long-horizon multimodal agency. →read the paper
| Research Highlights | 🌟🌟🌟 Almost everyone is talking about Google Research’s TurboQuant (though it’s not actually new, the initial paper was published in 2025 !). Why? TurboQuant is a compression algorithm for high-dimensional vectors. It combines PolarQuant, which rotates vectors and encodes them in polar coordinates to preserve signal with minimal overhead, with QJL, a 1-bit correction step that stores residual error as +1/−1 signs to keep scores accurate.
|  | Image credit: Google’s blog |
| The result: near-zero accuracy loss, up to 8× faster attention and vector search, roughly 6× smaller KV cache, and no retraining or fine-tuning. As long-context LLMs and billion-vector search grow, this kind of compression is becoming essential →read their blog | | | | ❓Important poll ❓ | Should we keep the Research Section? | | | Rest of research from the last week | (as always, 🌟 indicates papers that we recommend to pay attention to) | What we see this week: | Agents are becoming systems Memory is getting operational RL is going token-level Uncertainty is back Latency is a real bottleneck Multimodal agents are getting more active Video is becoming agentic Training is getting smarter Inference-time tricks are booming Specialized foundation models are expanding
| Agentic software engineering, repository memory, and execution harnesses | 🌟 Learning to Commit: Generating Organic Pull Requests via Online Repository Memory – Shows how coding agents can learn project-specific change patterns from a repository’s own history, which matters because maintainers reject a lot of machine-generated code for being alien to the codebase rather than plainly broken →read the paper 🌟 Effective Strategies for Asynchronous Software Engineering Agents – Demonstrates that multi-agent software work starts becoming useful when it borrows real software engineering primitives such as isolated branches, dependency-aware delegation, and structured integration instead of pretending coordination is magic →read the paper Natural-Language Agent Harnesses – Reframes harness engineering as a first-class artifact by moving agent control logic out of buried runtime code and into portable natural-language specifications, which is interesting if you care about comparability, reuse, and agent system design as an actual discipline →read the paper 🌟 Composer 2 Technical Report – Presents a strong example of what happens when a model is trained inside the same harness and tool structure it will actually use in deployment, making it relevant for anyone tracking the shift from benchmark performance to environment-grounded software agents →read the paper
| Agent workflows, optimization loops, and system-level orchestration | 🌟 From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents – Organizes the fast-growing workflow literature into a useful vocabulary around when structure is chosen, what gets optimized, and which signals guide changes, making it a strong map of the field rather than just a survey dump →read the paper Understanding the Challenges in Iterative Generative Optimization with LLMs – Exposes the hidden design choices inside self-improving agent loops, especially around what can be edited and what evidence gets fed back, which is exactly where many promising agent systems quietly fall apart →read the paper
| Efficient agent inference, routing, and latency reduction | SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning – Introduces a clever way to reduce agentic latency by using a lightweight model to speculate about tool-use trajectories before the expensive model finishes the whole loop, which matters because agent depth is becoming a real serving bottleneck →read the paper 🌟 Scalable Prompt Routing via Fine-Grained Latent Task Discovery – Tackles model routing at the level where the problem actually becomes hard: when many frontier models differ only in narrow capability pockets and broad taxonomies stop being useful →read the paper
| Training data mixtures and post-training efficiency | | Reinforcement learning for reasoning, with a focus on token-level mechanics | 🌟 On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation – Argues that the direction of RL-induced policy change matters more than raw magnitude, which is a sharper lens for understanding why RLVR works and how to exploit those updates both at training time and test time →read the paper 🌟 Sparse but Critical: A Token-Level Analysis of Distributional Shifts in RLVR Fine-Tuning of LLMs – Shows that RLVR improvements come from a surprisingly small set of token-level shifts, making the paper interesting as a mechanistic explanation of why reasoning gains can look broad while the actual policy changes stay narrow →read the paper 🌟 Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs? – Identifies a concrete failure mode in reasoning distillation by showing that compressing traces can suppress uncertainty expression, which is a useful corrective to the lazy assumption that shorter reasoning is always better reasoning →read the paper 🌟 Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models – Pushes reasoning beyond single-answer collapse by training models to generate multiple plausible answers with confidence structure in one pass, which is important for ambiguous, uncertain, or genuinely multimodal decision settings →read the paper
| Reinforcement learning and self-improvement for multimodal reasoning agents | Rethinking Token-Level Policy Optimization for Multimodal Chain-of-Thought – Brings token-level RL ideas into multimodal reasoning by separating perceptual grounding from exploratory inference, which is useful if you think multimodal CoT should not be optimized as one undifferentiated blob →read the paper When Models Judge Themselves: Unsupervised Self-Evolution for Multimodal Reasoning – Shows a path toward multimodal reasoning improvement without human-annotated answers or external reward models, which makes it relevant for scaling training when supervision becomes the bottleneck →read the paper
| Video agents and long-horizon audiovisual generation | EVA: Efficient Reinforcement Learning for End-to-End Video Agent – Treats video understanding as an active, query-driven process rather than passive frame ingestion, which is interesting because long video increasingly looks like an agent problem, not just a bigger context-window problem →read the paper Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models – Adapts RL to distilled autoregressive video models without re-distillation or reverse-process overhead, making it one of the more practically relevant attempts to align long-form video generation with human preferences →read the paper
| Vision representation upgrades and spatial understanding | SpatialBoost: Enhancing Visual Representation through Language-Guided Reasoning – Injects spatial knowledge into vision encoders through language-mediated reasoning, which is interesting because it uses linguistic structure to improve spatial representation instead of treating vision and language as cleanly separate stacks →read the paper MuRF: Unlocking the Multi-Scale Potential of Vision Foundation Models – Shows that a simple multi-resolution fusion trick can unlock better visual representations at inference time without retraining, which makes it broadly useful and pleasantly unglamorous in the best way →read the paper
| That’s all for today. Thank you for reading! Please send this newsletter to colleagues if it can help them enhance their understanding of AI and stay ahead of the curve. | Upgrade now | How did you like it? | | |
|
