Midnight reflexion on to improve scientific evaluation beyond h-index: how about a veting mechanism with pagerank based on e.g ORCID profiles. Each researcher can vouch for any number of other researchers. The importance of a researcher is then simply their pagerank.

Right now, SOTA models are all from AI labs. But suddenly all that investment in eg Mistral doesn’t look all that wasted, and for a non-US company it now looks strategic to be their customer (or a customer if any other non-US vendor) Needless to say huge advertisement for open models

The jailbreak that triggered all of this was reported by Amazon.

Actually curious how the translator would deal with this one. Any generous English speaking follow care sharing a screenshot (other languages are accepted too of course).

If you think a $300K corporate salary is payment for 40 hours of weekly labor, I've got news for you... There is a persistent cynical narrative that large enterprises are bloated engines of inefficiency, filled with overpaid professionals who spend their days looking at slides and doing "nothing." I mean, it's a comforting myth for critics, but I think it fundamentally misunderstands modern knowledge work. That $300K salary (or $400K, or $500K) isn't a reward for linear effort but an option premium on high-leverage thinking. We are still haunted by the ghost of the assembly line, ie, the outdated idea that compensation must directly correlate with time spent physical output. In the factory world, if you leave your station, production stops, but in the knowledge economy, value is almost totally decoupled from time. Folks... An enterprise paying a senior leader or specialist $25K a month is not buying 160 hours of typing, they are buying *insurance* against catastrophic errors and positioning themselves for asymmetric upside. I'll try to make it tangible with an example... Consider a complex matrix organization busy with a $40M product migration. In this environment, the value distribution of a worker's is heavily spiked. Most days look like nothing... alignment meetings, reading documentation, maintaining steady state. Yes, to an outsider, it looks like "doing nothing." But then a critical day arrives. A vendor fails, a timeline slips, a crossroads appears, whatever... If that $300K professional has the institutional memory and capability to make just 4 or 5 correct decisions during those critical moments, the ROI is staggering! A single right call can avert a $5M problem. Suddenly, that $300K salary doesn't look like bloat but, to me, seems like the cheapest asset on the p&l. These days we are bombarded by tech CEOs promising fully autonomous, AI-driven organizations and I keep saying these pitches miss the entire point of how complex enterprises actually move. Data computation can be outsourced to an LLM but going through the decision fabric of an enterprise cannot. You need people for: > Knowing *how* to build consensus across disconnected departments with competing incentives; > Understanding the unspoken history of why past projects failed, and how to position a new initiative so it doesn't trigger corporate antibodies; > When a multi-million-dollar decision goes sideways, an algorithm cannot stand before a board of directors or regulators and take ownership of the corrective action. An AI can give you a pristine strategic framework with nice and difficult sounding words, but it cannot navigate the human matrix required to execute it. The ability to be effective inside a complex enterprise is a rare AND expensive skillset precisely because it cannot be automated or easily replicated. My point is you aren't paying for the 9-to-5 "grind", but more for the readiness. Like an elite surgeon or an expert technician, you pay for the decades of accumulated knowledge that allow them to fix a crisis in 5 minutes, not the 5 minutes itself.... Leverage, not labor.

20/ Bottom line: these RNA localization elements are big, multipartite, and very complicated. We're digging further into the mechanisms behind their activity now. Stay tuned. Read the full preprint here 👇 biorxiv.org/content/10.64898…

Tl;de "make a copy of something that already exist" is now done extremely fast with Fable. Can't beat git pull yet but we might be getting there with enough compute.

It’s possible to deeply care about biosecurity and AI safety (hi that’s me) and also believe that keeping advanced models out of the hands of biologists is net negative for humanity (that’s also me)

Actually creepy. Who wants their AI assistant to profile them? They should just do the task as requested...

We need to continue to be honest about what our models are actually capable of and where they fail so we keep moving forward. Less hype & blind faith in parameters & arbitrary scaling. More evidence based strategies & honest reflection. 6/6

Love the simplicity. Powerful elegance.

When building and evaluating VCMs, we realized that at the end we always compare simulated results with pre-computed summary statistics. So… what if we used the observed summary statistics instead observed single cells for prediction?

These results demonstrate we can accomplish a lot by going back to basics and building models that, by design, reflect the statistics of the underlying data. Rhaister shows that scaling the right data is far more valuable than scaling parameters.

kinda funny that anthropic took a good hard look at the extreme nervousness that claude displays when answering questions as dangerous as ‘what is the powerhouse of the cell’ and decided they werent being conservative enough

Modern parental mode: the kids can't ask for some content *and* the assistant doesn't do anything unless they said please and will harass them until they say thank you.

2/ In hindsight, this is not a new observation. Parents have long complained that home assistants (Alexa/GA etc.) have made their children ruder IRL, since the devices don't require "please" and "thank you." What I see interacting with engineers using coding agents is an extrapolation of that.

Chers politicien.nes, Vous voulez réguler des cancérigènes ? Voici une liste : Le tabac L'alcool Les virus des hépatites et les HPV Les plages Les fast food Les sodas Les barbecue Les huiles essentielles. Ah et le tabac aussi. Et l’alcool.