The most recent trailer for the subsequent Star Wars movie is an explosive return to kind to the old style Saturday afternoon matinee type seen in classic Hollywood serials that includes Buck Rogers and Flash Gordon again within the heady days of pulp science fiction, one thing George Lucas’s authentic Star Wars Trilogy tried to faithfully seize and did so fantastically.
The surprisingly pleasant ultimate trailer for Disney/Lucasfilm’s The Mandalorian and Grogu was launched final week within the aftermath of CinemaCon 2026 and delivers greater than sufficient nostalgic house opera vibes previous to its scheduled Might 22, 2026 launch.
At that annual Las Vegas business conference, director John Favreau revealed the primary 17 minutes of the movie and wowed the attending viewers with thrills and spills from a galaxy far, distant, maybe reigniting that splendidly pure “Star Wars” magic.
Article continues under
Jabba the Hutt’s slimy warlord cousins generally known as “The Twins.” (Picture credit score: Disney/Lucasfilm)
In Disney/Lucasflim’s$166 million spinoff characteristic we discover Pedro Pascal reprising his function as feared bounty hunter Din Djarin from the Disney+ TV sequence, “The Mandalorian.” He is teaming up with the green-skinned, pint-sized Child Yoda to trace down wished criminals for the New Republic.
This cosmic journey is filled with Star Wars characters, ships and lore, together with a shiny new Razor Crest spaceship, hateful Hutt twins on the moon of Nar Shaddaa, mad mercenaries, mystical moments, monster battle droids, Imperial walkers, X-wings in formation, alien serpents, airborne Anzellans, and “Star Wars Rebels” fan favourite Zeb Orellios.
Two specific objects price mentioning are the humongous stop-motion robots Mando is combating which had been created by Academy Award-winning Phil Tippett and his ace VFX crew, and the extremely cool live-action rendering of Embo, “The Clone Wars”‘ brutal Kyuzo bounty hunter, and his devoted wolf-like creature Marrok. “The Mandalorian” co-creator Dave Filoni may even be offering Embo’s villainous voice!
Additionally starring Sigourney Weaver as New Republic Colonel Ward and Jeremy Allen White as Jabba’s son Rotta the Hutt, “The Mandalorian and Grogu” lands Might 22, 2026.
Breaking house information, the newest updates on rocket launches, skywatching occasions and extra!
Your brokers are solely nearly as good because the data they’ll entry — and solely as protected because the permissions they implement.
We’re launching ACL Hydration (entry management record hydration) to safe data workflows within the DataRobot Agent Workforce Platform: a unified framework for ingesting unstructured enterprise content material, preserving source-system entry controls, and implementing these permissions at question time — so your brokers retrieve the suitable info for the suitable person, each time.
The issue: enterprise data with out enterprise safety
Each group constructing agentic AI runs into the identical wall. Your brokers want entry to data locked inside SharePoint, Google Drive, Confluence, Jira, Slack, and dozens of different programs. However connecting to these programs is simply half the problem. The more durable downside is making certain that when an agent retrieves a doc to reply a query, it respects the identical permissions that govern who can see that doc within the supply system.
As we speak, most RAG implementations ignore this fully. Paperwork get chunked, embedded, and saved in a vector database with no file of who was — or wasn’t — purported to entry them. This may end up in a system the place a junior analyst’s question surfaces board-level monetary paperwork, or the place a contractor’s agent retrieves HR information meant just for inside management. The problem isn’t simply propagate permissions from the info sources throughout the inhabitants of the RAG system — these permissions should be constantly refreshed as persons are added to or faraway from entry teams. That is crucial to maintain synchronized controls over who can entry numerous varieties of supply content material.
This isn’t a theoretical threat. It’s the rationale safety groups block GenAI rollouts, compliance officers hesitate to log off, and promising agent pilots stall earlier than reaching manufacturing. Enterprise clients have been specific: with out access-control-aware retrieval, agentic AI can’t transfer past sandboxed experiments.
Present options don’t resolve this properly. Some can implement permissions — however solely inside their very own ecosystems. Others help connectors throughout platforms however lack native agent workflow integration. Vertical purposes are restricted to inside search with out platform extensibility. None of those choices give enterprises what they really want: a cross-platform, ACL-aware data layer purpose-built for agentic AI.
What DataRobot gives
DataRobot’s safe data workflows present three foundational, interlinked capabilities within the Agent Workforce Platform for safe data and context administration.
1. Enterprise knowledge connectors for unstructured content material
Hook up with the programs the place your group’s data truly lives. At launch, we’re offering production-grade connectors for SharePoint, Google Drive, Confluence, Jira, OneDrive, and Field — with Slack, GitHub, Salesforce, ServiceNow, Dropbox, Microsoft Groups, Gmail, and Outlook following in subsequent releases.
Every connector helps full historic backfill for preliminary ingestion and scheduled incremental syncs to maintain your vector databases present. You management entry and handle connections via APIs or the DataRobot UI.
These aren’t light-weight integrations. They’re constructed to deal with production-scale workloads — 100GB+ of unstructured knowledge — with sturdy error dealing with, retries, and sync standing monitoring.
2. ACL Hydration and metadata preservation
That is the core differentiator. When DataRobot ingests paperwork from a supply system, it doesn’t simply extract content material — it captures and preserves the entry management metadata (ACLs) that outline who can see every doc. Person permissions, group memberships, function assignments — all of it’s propagated to the vector database lookup in order that retrieval is conscious of the permissioning on the info being retrieved.
Right here’s the way it works (additionally illustrated in Determine 1 beneath):
Throughout ingestion, document-level ACL metadata — together with person, group, and function permissions — is extracted from the supply system and continued alongside the vectorized content material.
ACLs are saved in a centralized cache, decoupled from the vector database itself. It is a crucial architectural resolution: when permissions change within the supply system, we replace the ACL cache with out reindexing your entire VDB. Permission adjustments propagate to all downstream customers routinely. This consists of permissioning for domestically uploaded information, which respect DataRobot RBAC.
Close to real-time ACL refresh retains the system in sync with supply permissions. DataRobot constantly polls and refreshes ACLs inside minutes. When somebody’s entry is revoked in SharePoint or a Google Drive folder is restructured, these adjustments are mirrored in DataRobot on a scheduled foundation — making certain your brokers by no means serve stale permissions.
Exterior id decision maps customers and teams out of your enterprise listing (through LDAP/SAML) to the ACL metadata, so permission checks resolve appropriately no matter how identities are represented throughout completely different supply programs.
3. Dynamic permission enforcement at question time
Storing ACLs is critical however not adequate. The true work occurs at retrieval time.
When an agent queries the vector database on behalf of a person, DataRobot’s authorization layer evaluates the saved ACL metadata in opposition to the requesting person’s id, group memberships, and roles — in actual time. Solely embeddings the person is allowed to entry are returned. All the pieces else is filtered earlier than it ever reaches the LLM.
This implies two customers can ask the identical agent the identical query and obtain completely different solutions — not as a result of the agent is inconsistent, however as a result of it’s appropriately scoping its data to what every person is permitted to see.
For paperwork ingested with out exterior ACLs (akin to domestically uploaded information), DataRobot’s inside authorization system (AuthZ) handles entry management, making certain constant permission enforcement no matter how content material enters the platform.
The way it works: step-by-step
Step 1: Join your knowledge sources
Register your enterprise knowledge sources in DataRobot. Authenticate through OAuth, SAML, or service accounts relying on the supply system. Configure what to ingest — particular folders, file sorts, metadata filters. DataRobot handles the preliminary backfill of historic content material.
Step 2: Ingest content material with ACL metadata
As paperwork are ingested, DataRobot extracts content material for chunking and embedding whereas concurrently capturing document-level ACL metadata from the supply system. This metadata — together with person permissions, group memberships, and function assignments — is saved in a centralized ACL cache.
The content material flows via the usual RAG pipeline: OCR (if wanted), chunking, embedding, and storage in your vector database of alternative — whether or not DataRobot’s built-in FAISS-based answer or your individual Elastic, Pinecone, or Milvus occasion — with the ACLs following the info all through the workflow.
Step 3: Map exterior identities
DataRobot resolves person and group info. This mapping ensures that ACL permissions from supply programs — which can use completely different id representations — will be precisely evaluated in opposition to the person making a question.
Group memberships, together with exterior teams like Google Teams, are resolved and cached to help quick permission checks at retrieval time.
Step 4: Question with permission enforcement
When an agent or software queries the vector database, DataRobot’s AuthZ layer intercepts the request and evaluates it in opposition to the ACL cache. The system checks the requesting person’s id and group memberships in opposition to the saved permissions for every candidate embedding.
Solely licensed content material is returned to the LLM for response technology. Unauthorized embeddings are filtered silently — the agent responds as if the restricted content material doesn’t exist, stopping any info leakage.
Step 5: Monitor, audit, and govern
Each connector change, sync occasion, and ACL modification is logged for auditability. Directors can monitor who linked which knowledge sources, what knowledge was ingested, and what permissions had been utilized — offering full knowledge lineage and compliance traceability.
Permission adjustments in supply programs are propagated via scheduled ACL refreshes, and all downstream customers — throughout all VDBs constructed from that supply — are routinely up to date.
Why this issues to your brokers
Safe data workflows change what’s doable with agentic AI within the enterprise.
Brokers get the context they want with out compromising safety. By propagating ACLs, brokers have the context info they should get the job finished, whereas making certain the info accessed by brokers and finish customers honors the authentication and authorization privileges maintained within the enterprise. An agent doesn’t develop into a backdoor to enterprise info — whereas nonetheless having all of the enterprise context wanted to do its job.
Safety groups can approve manufacturing deployments. With source-system permissions enforced end-to-end, the chance of unauthorized knowledge publicity via GenAI isn’t simply mitigated — it’s eradicated. Each retrieval respects the identical entry boundaries that govern the supply system.
Builders can transfer sooner. As an alternative of constructing customized permission logic for each knowledge supply, builders get ACL-aware retrieval out of the field. Join a supply, ingest the content material, and the permissions include it. This removes weeks of customized safety engineering from each agent venture.
Finish customers can belief the system. When customers know that the agent solely surfaces info they’re licensed to see, adoption accelerates. Belief isn’t a characteristic you bolt on — it’s the results of an structure that enforces permissions by design.
Get began
Safe data workflows can be found now within the DataRobot Agent Workforce Platform. Should you’re constructing brokers that must cause over enterprise knowledge — and also you want these brokers to respect who can see what — that is the potential that makes it doable. Strive DataRobot or request a demo.
Google reportedly teased what’s arising for its Android Present: I/O 2026 Version in a now-removed YouTube video.
It has been acknowledged that the outline highlights the reveals because the “greatest” for Android but.
The Android Present is eyeing a Might 12 date at 10 am PT/12pm ET, per week earlier than I/O 2026 will reveal main Android 17 particulars and extra.
We’re already looking forward to I/O 2026 in Might; nevertheless, Google is seemingly teasing a present earlier than the principle occasion to get us prepared for the subsequent period of Android.
Google is reportedly teasing what’s arising for its Android Present, which can precede its I/O 2026 convention in Might (by way of 9to5Google). The video, which is not obtainable on YouTube on the time of writing, acknowledged the Android Present: I/O Version will start on Might 12 at 10 am PT (12 pm ET). There’s hypothesis that this present will contain consumer-oriented updates and insights for Android.
The publication highlighted the video’s description, which reportedly positions this 12 months’s present as probably the most vital one but. Google stated, “That is going to be one of many greatest years for Android but.” The temporary teaser concludes by stating, “be the primary to try what the longer term holds.”
Article continues under
There was fairly a bit lined final 12 months, so it is doubtless we’re taking a look at the same scenario.
It is the Android Present
(Picture credit score: Android)
If we have a look at what Google did final 12 months, its teaser for the Android Present (2025 version) featured its personal web site. The location, on the time, was fairly temporary, solely displaying the date, time, and a fast description of what customers may anticipate. It is just about no completely different than what Google’s achieved in the present day (or, not in the present day?) with its now-removed YouTube video. Final 12 months, the corporate’s Sameer Samat, President of Android Ecosystem, and members of the Android workforce, hosted the present.
The present usually airs per week earlier than I/O does, and it appears like we’re observing the same method this 12 months, too. The 2025 version held fairly a couple of hefty insights, similar to Materials 3 Expressive, Gemini‘s enlargement to extra units, and Gemini’s rip-off detection capabilities.
Following Google’s spherical of puzzles, customers revealed I/O 2026’s date, which is slated for Might 19-20. We additionally bought a good have a look at its periods record. Google is getting ready to speak about Android 17, its AI software program, and large Chrome updates for the 12 months.
Get the newest information from Android Central, your trusted companion on this planet of Android
Android Central’s Take
The Android Present is sort of a little house the place Google’s software program can shine slightly in the beginning else at I/O. What’s most curious is that the corporate teased main Android 17 highlights for I/O 2026, so what’s this present going to debate? Maybe any UI revamps will seem, alongside what’s coming for Gemini that issues on our telephones. The larger highlights will doubtless be reserved for the principle occasion the week after, however this may function an appetizer to maintain us going.
Get the Fashionable Science day by day publication💡
Breakthroughs, discoveries, and DIY ideas despatched six days per week.
Astronomers knew 3I/ATLAS wasn’t an area comet not lengthy after first recognizing it in July 2025. As solely the third interstellar object ever detected in our photo voltaic system, it supplied researchers a uncommon—and transient—alternative. With the proper timing and gear, scientists around the globe may study a cosmic customer who probably shaped below far totally different circumstances than these skilled in our personal area of the galaxy.
3I/ATLAS is now crusing away from Earth and our photo voltaic system itself, however astronomers have already discovered a wealth of data. The quickest comet ever recorded is roofed in ice volcanoes, and emits a dusty path of methanol and cyanide in its wake.
Earlier this month, the European House Company confirmed that 3I/ATLAS can be spewing the equal of 70 Olympic swimming swimming pools’ price of water daily. Nonetheless, the precise kind of water isn’t typically seen right here on Earth.In line with astronomers on the College of Michigan (UM), the hydrogen within the comet’s H2O incorporates one further neutron, which technically makes it an isotope known as deuterium. The rarity isn’t merely an odd quirk—it signifies 3I/ATLAS originated someplace a lot colder than the photo voltaic system.
“Our new observations present that the circumstances that led to the formation of our photo voltaic system are a lot totally different from how planetary programs advanced in numerous elements of our galaxy,” Luis Salazar Manzano, a UM astronomer, stated in an announcement.
The co-author of a paper printed at present within the journal Nature Astronomy, Manzano defined that 3I/ATLAS incorporates 30 occasions the deuterium seen in different comets, in addition to 40 occasions the quantity that exists in Earth’s oceans.
“The quantity of deuterium with respect to atypical hydrogen in water is increased than something we’ve seen earlier than in different planetary programs and planetary comets,” he added.
Measuring subatomic particles in a comet hundreds of thousands of miles away required a few of the most delicate instruments obtainable. Manzano and colleagues utilized gear on the MDM Observatory in Arizona, whereas additionally collaborating with astronomers on the Atacama Giant Millimeter/submillimeter Array (ALMA) in Chile. Because of ALMA, the crew may separate normal and deuterated water within the comet, then get correct ratio estimates between the 2. It’s not solely spectacular—it’s the primary time anybody efficiently achieved the evaluation on an interstellar object.
So what does a number of deuterium imply, precisely? For one factor, 3I/ATLAS’ birthplace was a lot colder than circumstances that created our photo voltaic system—lower than 30 levels Kelvin, or -387.67 Fahrenheit. The area doubtless additionally skilled a lot decrease ranges of radiation.
“Fuel-phase and ice-grain deuterium enrichments happen via chemical processes that function at low temperatures (<30 Ok) pointing in the direction of an origin within the prestellar molecular cloud or within the outer elements of the protoplanetary disk,” the examine’s authors wrote.
For the reason that Milky Means galaxy is an unlimited place, it might not come as an enormous shock to study different areas exhibit totally different formative environments. However as astronomer and examine co-author Teresa Paneque-Carreño defined, you may’t base science on assumptions—even after they sound ironclad on their very own.
“That is proof that regardless of the circumstances had been that led to the creation of our photo voltaic system are usually not ubiquitous all through house,” stated Paneque-Carreño. “Which will sound apparent, but it surely’s a kind of issues that it is advisable to show.”
2025 PopSci Better of What’s New
The 50 most necessary improvements of the 12 months
The Lacking White Home Tapes was a sketch comedy voice recording which
was a satiric commentary on the Watergate scandal. It was a spin-off
from Nationwide Lampoon journal. The recording was produced by Irving
Kirsch and Vic Dinnerstein. It was launched as a single on Blue Thumb
Information in 1973. In 1974 it was expanded into an album, which was
subsequently nominated for a Grammy Award as Greatest Comedy Recording of
the 12 months.
The only consisted of a doctored speech, wherein Richard Nixon
confesses culpability within the Watergate break-in. Aspect One of many album
accommodates extra doctored recordings of Nixon’s speeches and press
conferences. Aspect Two accommodates sketches carried out by John Belushi, Chevy
Chase, Rhonda Coullet, and Tony Scheuten.
I hadn’t considered, not to mention heard the document for years, however
just lately one thing (I do not keep in mind what) jogged my memory of it. A fast
Google search later…
I made some notes for a put up I’d really write up certainly one of lately
however, within the meantime, for these within the interval it is a
fascinating relic (to not point out a reminder that the Nationwide Lampoon
model was related to humor).
Apple’s product animations, significantly the scrolly teardowns (technical time period), have at all times been inspiring. However these bleeding-edge animations have at all times used JavaScript and different applied sciences. Plus, they aren’t at all times responsive (or, no less than, Apple switches to a static picture at a sure width).
I’ve been wowed by CSS’s newer scrolling animation capabilities and questioned if I might rebuild one in all these animations in simply CSS and make it responsive. (In reality, CSS certain has come a good distance since the final try on this publication.) The one I’ll be making an attempt is from the Imaginative and prescient Professional website and to see it you’ll must scroll down till you hit a black background, somewhat greater than midway down the web page. Should you’re too lazy errr… environment friendly to go look your self, and/or they resolve to alter the animation after this text goes dwell, you possibly can watch this video:
Notice: Whereas Apple’s model works in all main browsers, the CSS-only model, on the time of this writing, won’t work in Firefox.
Apple’s Animation
The very first thing we have now to do is determine what’s happening within the authentic animation. There are two main phases.
Stage 1: “Exploding” {Hardware}
Three digital parts rise in sequence from the Imaginative and prescient Professional machine on the backside of the web page. Every of the three parts is a set of two photographs that go each in entrance of and behind different parts like a sub roll round a sizzling canine bun round a bread stick. (Sure, that’s a bizarre analogy, however you get it, don’t you?)
The primary, outermost part (the sub roll) contains the frontmost and the hindmost photographs permitting it to look as if it’s each in entrance of and behind the opposite parts.
The subsequent part (the recent canine bun) wraps the third part (the bread stick) equally. This offers depth, visible curiosity, and a 3D impact, as clear areas in every picture enable the photographs behind it to indicate by.
Stage 2: Flip-As much as Eyepieces
The ultimate piece of the Imaginative and prescient Professional animation flips the machine up in a easy movement to indicate the eyepieces. Apple does this portion with a video, utilizing JavaScript to advance the video because the consumer scrolls.
Let’s recreate these, one stage at a time.
“Exploding” {Hardware}
Since Apple already created the six photographs for the parts, we will borrow them. Initially, I began with a stack of img tags in a div and used place: fastened to maintain the photographs on the backside of the web page and place: absolute to have them overlap one another. Nevertheless, once I did this, I bumped into two points: (1) It wasn’t responsive — shrinking the width of the viewport made the photographs go off display, and (2) the Imaginative and prescient Professional couldn’t scroll into view or scroll out of view because it does on the Apple website.
After banging my head towards this for a bit, I went again and checked out how Apple constructed it. That they had made every picture a background picture that was at background-position: backside middle, and used background-size: cowl to maintain it a constant facet ratio. I nonetheless wanted them to have the ability to overlap although, however I didn’t wish to pull them out of circulation the way in which place: absolute does so I set show:grid on their guardian ingredient and assigned all of them to the identical grid space.
.visionpro { /* the overarching div that holds all the photographs */
show: grid;
grid-template-columns: 1fr;
grid-template-rows: 1fr;
}
.half { /* every of the photographs has a component class */
grid-area: 1 / 1 / 2 / 2;
}
As my logic professor used to say within the early aughts, “Now we’re cooking with fuel!” (I don’t actually understand how that applies right here, nevertheless it appeared acceptable. Considerably illogical, I do know.)
I then started animating the parts. I began with a scroll timeline that might have allowed me to pin the animation timeline to scrolling your complete html ingredient, however realized that if the Imaginative and prescient Professional (that means the weather holding all the photographs) was going to scroll each into and out of the viewport, then I ought to change to a view timeline in order that scrolling the ingredient into view would begin the animation reasonably than attempting to estimate a keyframe proportion to begin on the place the weather can be in view (a reasonably brittle and non-responsive strategy to deal with it).
Scrolling the Imaginative and prescient Professional into view, pausing whereas it’s animating, after which scrolling it out of view is a textbook use of place: sticky. So I created a container div that totally encapsulated the Imaginative and prescient Professional div and set it to place: relative. I pushed the container div down previous the viewport with a high margin, and set high on the imaginative and prescient professional div to 0. You possibly can then scroll up until the place: sticky held the imaginative and prescient professional in place, the animation executed after which, when the container had been solely scrolled by, it could carry the Imaginative and prescient Professional div up and out of the viewport.
Now, to deal with the part strikes. Once I first used a translate to maneuver the photographs up, I had hoped to make use of the pure order of the weather to maintain all the pieces properly stacked in my bread-based turducken. Alas, the browser’s sneaky optimization engine positioned my sub roll solely on high of my sizzling canine bun, which was solely on high of my breadstick. Fortunately, utilizing z-index allowed me to separate the layers and get the overlap that’s a part of why Apple’s model seems to be so superior.
One other downside I bumped into was that, at sizes smaller than the 960-pixel width of the photographs, I couldn’t reliably and responsively transfer the parts up. They wanted to be far sufficient away that they didn’t intervene with Stage 2, however not so far-off that they went totally out of the viewport. (The place’s a bear household and a blonde lady once you want them?) Fortunately, because it so typically does, algebra saved my tuchus. Since I’ve the scale of the full-size picture (960px by 608px), and the total width of the picture is the same as the width of the viewport, I might write an equation like beneath to get the peak and use that in my translation calculations for a way far to maneuver every part.
Nevertheless, this calculation breaks down when the viewport is shorter than 608px and wider than 960px as a result of the width of the picture is now not equal to 100vw. I initially wrote an analogous equation to calculate the width:
However it additionally solely works if the peak is 608px or much less, and so they each gained’t work whereas the opposite one applies. This is able to be a easy repair utilizing an “if” assertion. Whereas CSS does have an if() operate as I’m penning this, it doesn’t work in Safari. Whereas I do know this complete factor gained’t work in Firefox, I didn’t wish to knock out a complete different browser if I might assist it. So, I fastened it with a media question:
I patted myself on the again for my mathematical genius and problem-solving abilities till I spotted (as you smarty pants individuals have most likely already discovered) that if the peak is lower than 608px, then it’s equal to 100vh. (Sure, vh is a sophisticated unit, significantly on iOS, however for this proof of idea I’m ignoring its downsides).
However no matter my mathematical tangents (Ha! Horrible math pun!), this allowed me to base my vertical translations on the peak of the Stage 2 graphics, e.g.:
…and thus get them out of the way in which for the Stage 2 animation. That mentioned, it wasn’t good, and at viewports narrower than 410px, I nonetheless needed to make an adjustment to the heights utilizing a media question.
Flip-As much as Eyepieces
Sadly, there’s no strategy to both begin a video with simply CSS or modify the body price with simply CSS. Nevertheless, we will create a set of keyframes that adjustments the background picture over time, similar to:
(Since there’s, like, 60-some photographs concerned on this one, I’m not supplying you with the total set of keyframes, however you possibly can go take a look at the cssvideo keyframes within the full CodePen for the total Monty.)
The draw back of this, nonetheless, is that as an alternative of 1 video file, we’re downloading 60+ information for a similar impact. You’ll discover that the file numbers skip a quantity between every iteration. This was me halving the variety of frames in order that we didn’t have 120+ photographs to obtain. (You would possibly be capable to pace issues up with a sprite, however since that is extra proof-of-concept than a production-ready resolution, I didn’t have the persistence to sew 60+ photographs collectively).
The animation was a bit uneven on the preliminary scroll, even when operating the demo regionally.
So I added:
…for each picture, together with the part photographs. That helped loads as a result of the server didn’t need to parse the CSS earlier than downloading all the photographs.
Utilizing the identical view timeline as we do for Stage 1, we run an animation shifting it into place and the cssvideo animation and the eyepieces seem to “flip up.”
Whereas a view timeline was nice, the animation didn’t at all times start or finish precisely once I wished it to. Enter animation-range. Whereas there’s a whole lot of choices what I used on all the .halfs was
animation-range: include cowl;
This made certain that the Imaginative and prescient Professional ingredient was contained in the viewport earlier than it began (include) and that it didn’t totally end the animation till it was out of view (cowl). This labored nicely for the components as a result of I wished them totally in view earlier than the parts began rising and since their endpoint isn’t vital they’ll hold shifting till they’re off display.
Nevertheless, for Stage 2, I wished to make sure the flip up animation had ended earlier than it went off display so for this one I used:
animation-range: cowl 10% include;
Each cowl and 10% discuss with the beginning of the animation, utilizing the cowl key phrase, however pushing its begin 10% later. The include ensures that the animation ends earlier than it begins going off display.
Right here’s all the pieces collectively:
And right here’s a video in case your browser doesn’t assist it but:
Conclusion
CSS certain has come a good distance and whereas I positively used some leading edge options there have been additionally a whole lot of comparatively current additions that made this doable too.
With scroll timelines, we will connect an animation to the scroll both of a complete ingredient or simply when a component is in view. The animation-range property allow us to fine-tune when the animation occurred. place: sticky lets us simply maintain one thing on display whereas we animate it at the same time as its scrolling. Grid format allowed overlap parts with out pulling them out of circulation. Even calc(), viewport models, customized properties, and media queries all had their roles in making this doable. And that doesn’t even depend the HTML improvements like preload. Unimaginable!
Possibly we should always add a W to WWW: The World Broad Wondrous Net. Okay, okay you possibly can cease groaning, however I’m not fallacious…
Apple is advancing AI and ML with basic analysis, a lot of which is shared by publications and engagement at conferences in an effort to speed up progress on this vital area and assist the broader group. This week, the Fourteenth Worldwide Convention on Studying Representations (ICLR) can be held in Rio de Janeiro, Brazil, and Apple is proud to once more take part on this vital occasion for the analysis group and to assist it with sponsorship.
A complete overview of Apple’s participation in and contributions to ICLR 2026 could be discovered right here, and a choice of highlights follows beneath.
Recurrent Neural Networks (RNNs) are naturally suited to environment friendly inference, requiring far much less reminiscence and compute than attention-based architectures, however the sequential nature of their computation has traditionally made it impractical to scale up RNNs to billions of parameters. A brand new development from Apple researchers makes RNN coaching dramatically extra environment friendly — enabling large-scale coaching for the primary time and widening the set of structure selections out there to practitioners in designing LLMs, notably for resource-constrained deployment.
In ParaRNN: Unlocking Parallel Coaching of Nonlinear RNNs for Giant Language Fashions, a brand new paper accepted to ICLR 2026 as an Oral, Apple researchers share a brand new framework for parallelized RNN coaching that achieves a 665× speedup over the standard sequential strategy (see Determine 1). This effectivity achieve permits the coaching of the primary 7-billion-parameter classical RNNs that may obtain language modeling efficiency aggressive with transformers (see Determine 2).
To speed up analysis in environment friendly sequence modeling and allow researchers and practitioners to discover new nonlinear RNN fashions at scale, the ParaRNN codebase has been launched as an open-source framework for computerized training-parallelization of nonlinear RNNs.
At ICLR, the paper’s first writer can even ship an Expo Discuss about this analysis.
Speedup from Parallel RNN Coaching
Determine 1: Runtime comparability for parallel and sequential software of the tailored ParaGRU and ParaLSTM cells as a perform of enter sequence size. ParaRNN unlocks training-time parallelizability, permitting dramatic speedups over vanilla sequential software.
Efficiency of Giant-Scale Basic RNNs
Determine 2: Perplexity (decrease is healthier) for varied mannequin sizes for Mamba2, ParaLSTM, ParaGRU, and a transformer. With large-scale coaching enabled by parallelization, the tailored GRU and LSTM fashions present perplexity aggressive with a transformer and Mamba2.
State House Fashions (SSMs) like Mamba have turn into the main different to Transformers for sequence modeling duties. Their major benefit is effectivity in long-context and long-form era, enabled by fixed-size reminiscence and linear scaling of computational complexity. To Infinity and Past: Device-Use Unlocks Size Generalization in State House Fashions, a brand new Apple paper accepted as an Oral at ICLR, explores the capabilities and limitations of SSMs for long-form era duties. The paper exhibits that the effectivity of SSMs comes at a value of inherent efficiency degradation. In actual fact, SSMs fail to unravel long-form era duties when the complexity of the duty will increase past the capability of the mannequin, even when the mannequin is allowed to generate chain-of-thought (CoT) of any size. This limitation arises from the bounded reminiscence of the mannequin, which limits the expressive energy when producing lengthy sequences.
The paper exhibits that this limitation could be mitigated by permitting SSMs interactive entry to exterior instruments. Given the proper selection of device entry and problem-dependent coaching knowledge, SSMs can be taught to unravel any tractable downside and generalize to arbitrary downside size and complexity (see Determine 3). The work demonstrates that tool-augmented SSMs obtain robust size generalization on a wide range of arithmetic, reasoning, and coding duties. These findings spotlight SSMs as a possible environment friendly different to Transformers in interactive tool-based and agentic settings.
Determine 3: Left: Illustration of an interactive tool-use agent trajectory with pointer-based reminiscence device for fixing multi-digit addition. The agent can generate ideas (blue), outputs (purple) or instructions (orange), and obtain observations (inexperienced) from the reminiscence device. At every step, we present the state of the reminiscence context on the highest row, and beneath it present the sequence of generated tokens. Proper: Accuracy of recurrent/SSM fashions (Mamba, LSTM, GRU) and Transformers (Pythia, Mistral) educated on trajectories for ≤5-digit addition, evaluated on as much as 1,000-digits (log scale).
Unified multimodal LLMs that may each perceive and generate photos are interesting not just for architectural simplicity and effectivity, but in addition as a result of shared representations may end up in deeper understanding and higher vision-language alignment, and might allow distinctive capabilities like picture enhancing by directions.
Nevertheless, present open-source fashions usually undergo from a efficiency trade-off between picture understanding and era capabilities. At ICLR, Apple researchers will share MANZANO: A Easy and Scalable Unified Multimodal Mannequin with a Hybrid Imaginative and prescient Tokenizer. As described within the paper, Manzano is a unified framework designed to scale back this efficiency trade-off with a easy architectural concept (see Determine 4) and a coaching recipe that scales effectively throughout mannequin sizes.
Manzano makes use of a single shared imaginative and prescient encoder to feed two light-weight adapters that produce steady embeddings for image-to-text understanding and discrete tokens for text-to-image era inside a shared semantic area. A unified autoregressive LLM predicts high-level semantics within the type of textual content and picture tokens, and an auxiliary diffusion decoder then interprets the picture tokens into pixels. This structure, along with a unified coaching recipe over understanding and era knowledge, permits scalable joint studying of each capabilities. Manzano achieves state-of-the-art outcomes amongst unified fashions, and is aggressive with specialist fashions, notably on text-rich analysis.
Determine 4: Our hybrid tokenizer workflow. (Left): The tokenizer produces two distinct however homogeneous characteristic streams by separate adapters. Throughout coaching, one adapter output is randomly sampled and handed to a small LLM decoder for alignment. (Proper): As soon as the tokenizer is educated, the proper panel illustrates how these two characteristic sorts are utilized to understanding and era duties.
At ICLR, Apple researchers can even share Sharp Monocular View Synthesis in Much less Than a Second, which presents a technique for producing a 3D Gaussian illustration from {a photograph}, utilizing a single ahead go by a neural community in lower than a second on a normal GPU. The ensuing illustration can then be rendered in actual time from close by views, as a high-resolution photorealistic 3D scene (see Determine 5).
Known as SHARP (Single-image Excessive-Accuracy Actual-time Parallax), this method delivers a illustration that’s metric, with absolute scale, supporting metric digicam actions. Experimental outcomes display that SHARP delivers strong zero-shot generalization throughout datasets. It additionally units a brand new cutting-edge on a number of datasets, lowering LPIPS by 25-34% and DISTS by 21-43% versus one of the best prior mannequin, whereas decreasing the synthesis time by three orders of magnitude.
To allow the group to additional discover and construct on this strategy, code is on the market right here.
ICLR attendees will have the ability to expertise this work firsthand in a demo on the Apple sales space #204 throughout exhibition hours.
Determine 5: SHARP synthesizes a photorealistic 3D illustration from a single {photograph} in lower than a second. High: Enter picture; Backside: Novel view synthesized by SHARP. The synthesized illustration helps high-resolution rendering of close by views, with sharp particulars and high quality constructions, at greater than 100 frames per second on a normal GPU.
Protein folding is a foundational but notoriously difficult downside in computational biology. At its core, this downside includes predicting the exact three-dimensional coordinates for every atom inside a protein construction, based mostly solely on its amino acid sequence (i.e., a string of characters with 20 potential values for every character). Predicting the 3D construction of proteins is critically vital as a result of a protein’s perform is inherently linked to its spatial configuration. Breakthroughs on this space allow researchers to quickly design and perceive proteins, probably revolutionizing drug discovery, biotechnology, and past.
At ICLR, Apple researchers will share SimpleFold: Folding Proteins is Easier than You Suppose, which particulars a brand new strategy that makes use of a general-purpose structure based mostly solely on normal transformer blocks (much like text-to-image or text-to-3D fashions). This strategy permits SimpleFold to dispense with the advanced architectural designs of prior approaches, whereas sustaining efficiency (see Determine 6). To allow the analysis group to construct on this technique, the paper is accompanied by code and mannequin checkpoints that may be effectively run regionally on Mac with Apple silicon utilizing MLX.
Determine 6: Instance predictions of SimpleFold on targets (a) chain A of 7QSW (RubisCO massive subunit) and (b) chain A of 8DAY (Dimethylallyltryptophan synthase 1), with floor reality proven in mild aqua and prediction in deep teal. (c) Generated ensembles of goal chain B of 6NDW (Flagellar hook protein FlgE) with SimpleFold fine-tuned on MD ensemble knowledge. (d) Efficiency of SimpleFold on CASP14 with growing mannequin sizes from 100M to 3B. (e) Inference time of various sizes of SimpleFold on an M2 Max 64GB MacBook Professional.
Throughout exhibition hours, ICLR attendees will have the ability to work together with reside demos of Apple ML analysis in sales space #204 together with:
SHARP – This demo exhibits SHARP working on a set of pre-recorded photos or photos captured straight by the consumer throughout the demo. Guests will expertise the quick course of from deciding on a picture, processing it with SHARP, and viewing the generated 3D Gaussian level cloud on iPad Professional with the M5 chip.
Native LLM inference on Apple silicon with MLX – This demo will showcase on-device LLM inference on a MacBook Professional with M5 Max utilizing MLX, Apple’s open-source array framework purpose-built for Apple silicon, working a quantized frontier coding mannequin fully regionally inside Xcode’s native improvement setting. The complete stack — MLX, mlx-lm, and mannequin weights — is open supply, inviting the analysis group to construct on and lengthen these strategies independently.
We’re proud to once more sponsor affinity teams internet hosting occasions onsite at ICLR, together with Girls in Machine Studying (WiML) (social on April 24), and Queer in AI (social on April 25). Along with supporting these teams with sponsorship, Apple workers can even be taking part in these and different affinity occasions.
ICLR brings collectively professionals devoted to the development of deep studying, and Apple is proud to once more share revolutionary new analysis on the occasion and join with the group attending it. This submit highlights only a choice of the works Apple ML researchers will current at ICLR 2026, and a complete overview and schedule of our participation could be discovered right here.
We ran a two-week proof of idea. I cut up the cluster into two swimming pools: Eight GPUs devoted to immediate processing and the remaining GPUs dealing with token era. No new {hardware}, no new cluster — only a configuration change within the serving layer and a routing coverage that despatched every request to the suitable pool primarily based on its inference section. The prompt-processing pool hit 90–95% compute utilization persistently as a result of that’s all it did. No token era competing for scheduling slots. No decode requests sitting idle whereas a prefill burst hogged the cores.
The token-generation pool was the larger shock. By batching tons of of concurrent decode requests collectively the reminiscence reads obtained amortized throughout extra work. Bandwidth utilization climbed above 70% — much better than the 30% we’d been seeing when decode requests had been interleaved with prefill on the identical GPU. General compute effectivity roughly doubled.
The associated fee math adopted. The client was spending about $2M yearly on inference GPU-hours. After disaggregation they had been on monitor to chop that by $600–800K whereas serving the identical request quantity on the identical latency targets. No new {hardware} bought. Identical GPUs, identical cluster, identical mannequin weights — completely different structure.
Nicer climate provides you one much less excuse to not be chasing your health objectives. If music or podcasts additionally assist you to push via the burn, some health earbuds may be a very good match, and immediately’s deal is a good probability to choose some up at a discount. The JBL Endurance Race 2 earbuds are an excellent selection for anybody seeking to get pleasure from music whereas staying energetic, and so they simply dropped to $42.46 on Amazon for the primary time — 47% off the conventional value of $79.95.
We had been alerted to this supply by way of our companion AI-powered value intelligence platform that tracks hundreds of merchandise, analyzes historic pricing knowledge, and helps you make knowledgeable buying selections. AI could make errors, so it’s at all times price checking the deal for your self, however this looks as if a real all-time low value. Beneath, you may see some useful AI insights in regards to the deal that will help you determine if you wish to make the most of the supply.
These true wi-fi earbuds function an IP68 waterproof and dustproof design, making them good for intense exercises and even outside adventures. With 48 hours of whole playback time, you may hear all day with out worrying about recharging. The 6.8mm dynamic drivers ship highly effective JBL Pure Bass sound, whereas 4 microphones guarantee clear calls, even in noisy environments. Plus, Energetic Noise Canceling helps you focus by filtering out distractions.
In response to our AI companion web site, which tracks Amazon costs and charges offers utilizing AI evaluation, this the Endurance Race 2 scores 97 out of 100, making it a superb deal. This excessive rating displays a powerful value benefit, with the present value being $25.01 under the 90-day common of $67.47. The earbuds are at an all-time low, and the value was simply dropped 23 hours in the past.
Thanks for being a part of our neighborhood. Learn our Remark Coverage earlier than posting.
The bottom of the lamp has two slider buttons. One toggle adjusts the heat, from chilly white mild all the way in which to crimson. One adjusts the depth, from ultra-bright right down to a glareless glow. Arduous faucets on every button skip forward, whereas holding the toggle down on one aspect or one other adjusts the sunshine settings fairly slowly—slowly sufficient I at first generally query whether or not it is occurring.
The utmost brightness is 1,000 lumens—the approximate depth of a 75-watt incandescent bulb. At this brightness, the battery lasts about 5 hours. At a decrease depth, this could prolong to so long as a dozen hours.
Purple mild remedy is, after all, the province of TikTok as a lot as science—a subject the place wild exaggerations dwell alongside authentic makes use of and advantages. For each sleep research exhibiting that crimson mild is superior to blue mild with regards to melatonin ranges, there’s one other exhibiting that crimson mild is related to “destructive feelings” earlier than mattress.
So I can solely supply my very own expertise, which is that Edge Gentle Go’s crimson studying mild affords me a nice liminal area between awake time and sleepy time, one not supplied by a fundamental nightstand lamp. It permits me to form of delight in a darkroom area that also lets me see and browse, and drift off just a little simpler.
If I go to sleep, the sunshine has an automated 25-minute shut-off, which suggests I will not do what I far too typically do, which is drift off whereas studying after which get up, alarmed, to a room crammed with brilliant mild in the midst of the evening.
Caveats and Quirks
{Photograph}: Matthew Korfhage
This mentioned, for all of the virtues of portability, the Edge Gentle Go doesn’t boast a base that is heavy sufficient to cease the lamp from tipping over if I bend it ahead from its lowest hinge. This may be an annoyance when making an attempt to make use of the lamp as a studying mild from a bedside desk or the arm of a sofa.