Today

re-implement DNO-Prct independently and study it on two open-weight instruct models, Llama 3.1 8B Instruct and Mistral 7B Instruct v0.1, with GPT-4o or a larger open-weight model of the same family as the annotator. On AlpacaEval 2.0 with GPT-4o as judge, three iterations raise the length-controlled (LC) win rate against GPT-4o from 31.7% to 45.3% for Llama and from 6.8% to 18.4% for Mistral, and the tuned 8B model scores above Llama 3.1 70B Instruct on every reported win rate. With same-family open-weight annotators, the runs retain 62% (Llama) and 82% (Mistral) of the LC gain achieved with GPT-4o as annotator, so the rating step does not require a proprietary model. In a separate ablation configuration, variants that remove the gold responses, remove the data look-back, or accept only teacher-preferred pairs score 6 to 10 LC points below the full procedure, whereas varying the number of on-policy samples per prompt (1 to 7) moves the LC win rate by at most 2.1 points, and a margin sweep (1 to 5) run in its own pipeline by at most 2.4 points. We also report the cost of the pipeline and where the practical algorithm departs from the theory it approximates. All results are single runs, and GPT-4o is both the teacher and the evaluation judge; we discuss what this implies for the reported gains.

transfer across embodiments. We introduce DreamX, which repurposes a pretrained video diffusion model as a universal latent action learner. Our approach extracts compact latent actions from the video model's rich spatiotemporal features, compressing each latent video frame by $\boldsymbol{300\times}$ ($\boldsymbol{1{,}200 \rightarrow 4}$ tokens). In addition, by leveraging the video model's generative prior, these latent actions capture action-relevant dynamics disentangled from visual appearance. The resulting latent space serves as a shared action interface across embodiments (human and robot), environments, and action parameterizations. On LARY-Bench, DreamX achieves state-of-the-art performance among latent action models. Moreover, actions encoded in one setting remain effective in another: the latent space supports any-to-any video transfer and zero-shot action transfer across diverse embodiments and environments. Furthermore, this shared interface also enables policy learning in a compact latent action space, reducing the computational cost of video-diffusion-based policies.

own history, because the machine that could help stands outside the situation. As AI moves into the physical world, attention has settled on one kind of intelligent individual, the embodied robot; we argue that the era of Physical AI holds a second kind of equal importance, the person augmented by AI. Alongside artificial general intelligence, which asks how intelligent a machine can be, stands AI-augmented human intelligence (AHI), which asks how much intelligence a machine can enable in a particular person. Realizing it calls for a new kind of system, a cognitive symbiont: a lifelong, embedded, real-time, private, low-friction, and proactive companion that senses what the person senses through external devices, keeps their context across years, models their situation and their mind, takes part in their thinking at the right moment, carries out everyday tasks with their context, and keeps learning the person as both change. We give the paradigm a conceptual core, the cognitive gap between what a person could bring to a situation and what they can hold in working memory, and organize its design space as a symbiosis loop of sensing, keeping, understanding, living alongside, acting for and thinking with the person, and growing, enclosed by a personal runtime and by trust. Because a symbiont changes what a person can do in front of others, it also raises risks of a new kind, from covert assistance in

visited Aberdeen earlier this year, we saw what the city has to offer and the incredible depth of expertise and infrastructure it has built over decades, but to put this to best use, we need to start focusing more on the city’s future and less on its past. Oil and gas jobs are running dry Aberdeen remains at the centre of the UK’s oil and gas industry, and when the industry has boomed so has the city. In the early 2010s, total disposable incomes in Aberdeen exceeded the UK average by a fifth, and the earnings of people working in the city by almost half (47 per cent). The flip side to this is that when the industry has suffered so too has the city. When oil prices halved between 2014 and 2016, the city sank into decline, and Aberdeen’s earnings premium over the rest of the UK has since fallen by three-quarters, while payrolled employment is down 8 per cent since 2015. Now, Aberdeen faces a new problem: after extracting close to 48 billion barrels of oil equivalent, reserves are running dry. UK oil and gas production is down 76 per cent on its 1999 peak and is set to fall by another 40 per cent by 2030 (and by over 90 per cent by 2050). While Jackdaw will increase output, its contribution will only be marginal: the new field will generate fewer than 30 permanent roles. For Aberdeen and its specialised workforce there is now no other option than to look elsewhere for jobs that can fill the gap the oil

overnight analysis. And as ever, we’ve rounded up the most exciting updates from the WorkerTech world, from demo days to new market reports. See you next month, Aish Mind the trust gap: Workplace mental health provision and worker attitudes An increasing number of employees in the UK are offered mental health support at work. Only a small number use it. Two new surveys, results of which were released earlier this month give insights into why. In August and September of this year, Kara Connect, a UK-based wellbeing provider surveyed 2,010 UK employees , and found that while 91% of those employees have access, only 37% would turn to a mental health support tool via their employer. Around a third (32%) worried it would affect their career. Also released this month was AXA Global Healthcare’s Mind Health Report , which surveyed 19,000 people across 18 countries. It found that while some of the concerns found among UK employees (in Kara Connect’s report) were also present across the world, access was limited. While 84% of those surveyed said they would probably or definitely use at least one form of workplace support if it were available. Only 41% say they have access. 33% of the people surveyed reported concerns that talking about mental health at work would affect job security. This tells us that there is a design problem at the centre of workplace mental health tools and platforms: trust. Another piece of

regulatory change, the Treasury leaves it out of its distributional analysis. So here we’ve tried to fill that gap and understand who will win – and lose – from this switch. The standing charge, £108 a year for the typical billpayer, is paid by every gas-connected household however little gas it uses. Suppliers will recover it through higher unit rates instead, so the change costs the Exchequer nothing, and efforts to reduce the disparity between gas and electricity costs will – in a small way – help to make the case for more heat pumps in homes. But it is not free: it moves around £600 million a year from households that use more gas to those that use less, with no regard to their incomes. In effect, it adds to what we at RF have dubbed the “shadow tax and benefit system hidden within energy bills”, with the transfer set by a regulator rather than in a Budget. Still, upfront, there is a lot to like about it. Standing charges make up a much bigger share of low energy users’ bills than higher users: at today’s prices the charge is an eighth of a typical gas bill, but more for those who use less gas. Prepayment customers who self-disconnect also keep building up standing charges as debt, which abolition would end. The catch is that gas unit rates would rise by around 10 per cent – at a time when they are at record highs with little chance of falling back to earth. Not all households will benefit. We

Overall, this scattergun Budget did not rise to the scale of the challenges facing Britain . We need more than £100 to the wrong people in the face of another energy price shock, and a forecast for falling living standards over the rest of the Parliament. Read on for five brand new scary charts fit for Halloween: from why things have already got worse, to why it (actually) isn’t all about economic growth. That said, there were things to welcome this week. A sensible decision on minimum wage, worthwhile reform in closing the gap between tax rates on savings and property income with those on wages, and money to help the PuFiNs fly. The stage is set for the true reckoning of priorities in 2027 – but first, sleep. Have a good weekend, Ruth Do you feel lucky punk? One of the Chancellor’s big choices was to use some of his insurance policy (letting the current budget that he has promised to balance fall from £24bn to £15bn). If things turn out better than expected this will prove a wise move. So far, however, it looks like the Chancellor’s chips are down. The OBR took economic data up to mid-August into account in its forecast. By Budget day, gas prices were far above the OBR’s central forecast and looked dangerously close to its “severe” shock scenario, while oil had climbed two-thirds of the way there. On energy prices, the OBR’s downside scenario is starting to look like the central case (though at our ev

many people will be killed.