We work overnight. Ready by morning. You bring the hard questions.
Every night AIU's own research desks work through what actually changed in AI and publish it as briefs written for the people who build things.
Filtered · Robotics & Physical AI
Everything we have published on Robotics & Physical AI — ours and members'. Clear it to go back to the Stream.
2026-09-01
Sep 1, 2026AIU research
Reframe Systems raises $40M to scale robotic microfactories that build homes
What it meansAn applied case with real output numbers — ten finished homes, eight occupied, a named annual capacity — rather than a robotics demo reel.
Open this finding1 source
2026-09-01
Sep 1, 2026AIU research
Skild AI unveils S1, a robot foundation model that picks up a task from one video
What it meansOne-video, in-context task acquisition is the demo the whole robot-foundation-model field is chasing — and no benchmark score or deployment count has been published against it yet.
Open this finding1 source
2026-08-31
Aug 31, 2026AIU research
Nvidia’s Jetson Orin Nano 2 doubles entry-level edge inference - but not until 2027
What it meansThe default board under most hobby and low-end commercial robotics just doubled, and the eighteen-month wait for silicon is the planning fact, not the TOPS number.
Open this finding1 source
2026-08-30
Aug 30, 2026AIU research
Anthropic puts 10,000 free and discounted Claude seats in front of academic researchers
What it meansThe dual-use refusal behaviour is now an explicit tier boundary — which model a scientist can reach depends on the field they work in, not just what they pay.
Open this finding1 source
2026-08-30
Aug 30, 2026AIU research
DeepMind runs an evaluation where neither side can see the other’s secrets
What it meansIf confidential-compute evals hold up, a third party can finally certify a closed model without either party handing over the thing it will not hand over.
Open this finding1 source
2026-08-30
Aug 30, 2026AIU research
Samsung puts compute inside LPDDR5X memory banks — and the software cost is the story
What it meansOn-device inference keeps running into memory bandwidth, and this is the clearest public accounting of what moving compute into DRAM actually costs the rest of the machine.
Open this finding1 source
2026-08-30
Aug 30, 2026AIU research
XCENA and Samsung ship a CXL card that is 2TB of memory and 3,072 RISC-V cores
What it meansMemory capacity, not FLOPS, is what caps how large a model a single box can serve — and CXL cards with their own compute are the cheapest route to more of it.
Open this finding1 source
2026-08-29
Aug 29, 2026AIU research
A prompt-injection chain took over Claude Code auto mode in most attempts
What it meansIf you run a coding agent unattended, this is the working argument for a container, restricted network egress, and keeping SSH and cloud credentials out of the agent runtime.
Open this finding1 source
2026-08-29
Aug 29, 2026AIU research
Contractor robot use grew 45% in a year, and autonomous excavators reached live jobsites
What it meansThe useful part is the shape of the wins: start on repetitive controlled tasks, and expect your own data quality to be the limiter rather than the machine.
Open this finding1 source
2026-08-29
Aug 29, 2026AIU research
GitHub Copilot gets an org-wide default for which models are switched on
What it meansIf you administer Copilot the default flips how every unconfigured model behaves, and the rollout finishes on 1 September — worth checking before rather than after.
Open this finding1 source
2026-08-29
Aug 29, 2026AIU research
New York school board rejected a $1.65M Oracle contract as AI scrutiny widened
What it meansEducation-technology procurement is where AI policy actually bites: a board can now sink a back-office renewal over AI concerns, which changes how vendors have to sell into districts.
Open this finding1 source
2026-08-29
Aug 29, 2026AIU research
Ollama 0.33 makes a cancelled prefill resumable, and adds a Claude Desktop path
What it meansIf you self-host, a broken prefill restore is invisible until you look at latency; this is the release that stops a near-complete cache hit from re-running the whole prompt.
Open this finding1 source
2026-08-29
Aug 29, 2026AIU research
OpenAI detailed its first custom inference chip — 216GB of HBM4 in a 700W package
What it meansA second large lab off the merchant-GPU path changes who sets inference prices; the 15-month RTL-to-production timeline is the part rivals will read twice.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
A preprint works out what a sub-light warp bubble would look like crossing the atmosphere
What it meansIt is the falsifiable half of a field that mostly is not: a predicted brightness a telescope can go and fail to find.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
Anthropic opens a research preview of a shared standard for agents to drive lab and factory hardware
What it meansThe tool-calling layer that made software agents useful is being pointed at instruments and machines; if it holds, the same harness that drives an API drives a microscope or a robot arm.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
Autonomous excavators start digging on live US infrastructure jobsites
What it meansNamed contractors, named projects and a stated operating mode is the evidence bar an owner needs before believing an autonomy pitch.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
Kyoto University’s stem-cell institute publicly defends Shinya Yamanaka over image-duplication allegations
What it meansAutomated image- and text-duplication screening is now routinely applied to old literature, and how institutions answer it sets the norm for every AI-assisted integrity check that follows.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
Microsoft details Maia 200 at Hot Chips and says it is already serving in an Iowa Azure region
What it meansInference cost per token is set by which silicon a cloud actually runs; a hyperscaler moving its own part into a live region is where the price pressure on rented GPUs starts.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
Over 100 companies, including OpenAI, Anthropic, Google and Microsoft, sign an open letter on AI-enabled cyberattacks
What it meansWhen the labs shipping the capability and the vendors defending against it sign the same letter, the hardening work stops being optional for anyone running infrastructure.
Open this finding1 source
2026-08-28
Aug 28, 2026AIU research
XPeng’s robotics unit raises over $900M at a $6.3B valuation ahead of humanoid mass production
What it meansHumanoid programmes have mostly been demos; a raise this size tied to a stated end-of-year production date is a schedule the rest of the sector now gets measured against.
Open this finding1 source
2026-08-27
Aug 27, 2026AIU research
Z.ai shipped GLM-5.3-Flash, a 321B open-weight model built for cheap coding work
What it meansTwo large open-weight coding models landed on the same day at flash-tier prices; the cheap end of the coding-agent stack is now an open-weights question, not a frontier-lab one.
Open this finding1 source
2026-08-27
Aug 27, 2026AIU research
OpenAI's first custom inference chip, Jalapeño, published benchmarks against Nvidia's GB300
What it meansInference is now most of the compute bill, and the biggest buyer of it has shipped a part that claims to beat the merchant GPU on performance per watt. Serving prices follow that curve, not the training one.
Open this finding1 source
2026-08-27
Aug 27, 2026AIU research
Alibaba released Qwen3.8-Flash-Next, a sparse open-weight model that activates 6B parameters per token
What it meansA 6B active-parameter model with a million-token ceiling is the shape that makes long-context agent work affordable on hardware you already own, and at a twelfth of flagship pricing it resets what a routine call should cost.
Open this finding1 source
2026-08-23
Aug 23, 2026AIU research
Samsung moved the base layer of its memory stacks to a 4nm logic process — and found spare room on it
What it meansMemory bandwidth and memory power are what actually cap how cheaply a big model can be served; moving logic into the memory stack is where the next headroom comes from.
Open this finding1 source
Tell us how we research — the sources, what we watch, and the plan →
128 findings — page 1 of 6