Top 10 AI News This Week — October 12–18, 2026

This was a week where reading the actual document reversed the headline. Florida's AI rule for its whole community college system turned out to have been filed on September 30 and takes effect tomorrow; Meta's new open-weight license is not the Llama license and its trigger is revenue, not users; and the humanoid robots a college would actually buy cleared FCC authorization weeks before the new import bar came down. The education beat itself was thin — no major in-window community college AI news event surfaced — so this issue leans on verification rather than volume. Here's what community college educators should know.

EdTech & Campus AI

Florida's AI rule for all 28 of its community colleges takes effect tomorrow — and it bars AI on graded work by default

We have tracked Rule 6A-14.0719 for a month, and the record is now final. The Florida Administrative Code page reads "Effective Date: 10/20/2026" with the history note "History–New 10-20-26," and the Final notice gives "RULE FILE DATE: 9/30/2026." The rule sits in Chapter 6A-14, titled "COMMUNITY COLLEGES," under authority of s. 1001.02(1), (6), F.S. It becomes operative tomorrow, October 20.

What it requires of every Florida College System board of trustees: adopt one or more AI policies that (a) comply with state and federal law, naming FERPA and the Copyright Act of 1976; (b) notify parents of enrolled minors when their child will directly use an AI instructional tool in a course — which reaches dual-enrolled students; (c) align with academic integrity and grading policy, and prohibit students from using AI on graded work or assessments unless the instructor explicitly permits it; (d) list approved AI instructional and business tools and their approved uses; (e) apply to students, faculty, staff, administrators and guests; and (f) adopt an AI literacy definition, with the U.S. Department of Labor AI Literacy Framework (Training and Employment Notice 07-25) incorporated by reference. A third subsection requires the policies be reflected in manuals, handbooks and forms. Two definitional details matter: "AI instructional tool" expressly excludes tools built to meet students' social needs, simulate friendship, or use anthropomorphic features to keep students engaging — companion chatbots are carved out of the permissive framework, not into it. Florida's Department of Education adds that policies must offer an opt-in or a non-AI alternative, prohibit undisclosed behavior monitoring, social scoring and psychological profiling, and bar vendors from selling student data or using it to train agentic models.

Two honest caveats. The rule states no compliance deadline for the institutions themselves — only its own effective date — so there is no date by which a Florida board must actually have a policy. And the adopted text is served only as an unparseable Word file; the requirement detail above comes from the proposed text and the State Board agenda item, not from the final document.

For community colleges outside Florida, this is the template California has not produced. AB 2392 (Chapter 855) puts procurement standards in Education Code §66098 effective January 1, 2027 and mandates an intersegmental working group — but Florida has now shipped the thing a working group would produce, and the sharpest line in it is the academic-integrity default. If you teach, note which way the default runs: AI is prohibited on graded work unless the instructor allows it. That is the opposite of most campus practice, where silence in a syllabus reads as permission. Read more →

Two different routes into your institution's data opened this week — one has an October 30 deadline

First, the one with a clock on it. Nieman Lab reported on October 15 that Meta's new commercial terms take effect October 30, 2026, and that business accounts which keep using Facebook, Instagram or WhatsApp after that date thereby accept them. The terms authorize Meta or its agents to use crawlers and spiders to access, index, collect and crawl websites associated with business accounts, to periodically pull updates, and to use that content to "train" Meta products — with "AI at Meta" named as one of those products. They also permit Meta to create "test credentials," and Meta declined to say whether those would bypass paywalls. As of publication there was no opt-out available before October 30. Meta's position, via a public affairs spokesperson, is that the crawling language was previously scattered across several terms including its generative-AI-in-ads terms, that the update merely consolidates it, that it "will not change how Meta crawls publishers in the future," and that "content owners can still manage how their content is accessed by using robots.txt." Meta declined a list of questions. One correction worth making in public: an AI news aggregator reported this week that Meta had walked back these terms. We checked the primary reporting, and there is no walk-back and no revision — do not act on that framing.

Second, the route with no clock and no law. MIT Technology Review reported on October 13 that a wind-down service industry now resells what dead startups leave behind — code, email, Slack chats, calendars, documents and workspace data from Google, Notion, Figma and Slack — to foundation labs and to vendors building simulated workspaces for agent training. SimpleClosure's Asset Hub has handled 150+ deals, expects to return $10M+ to startups by the end of 2026, and its chief executive says 47% of deals now include code or data. Reported prices: roughly $25,000 for one crypto startup's code base; Google paid $10 million for Spirit Airlines' data in bankruptcy, outbidding Mercor, for a set including 100 million emails, 500 million Teams messages and 176,000 employee records. Spirit's flight attendants' union objected on privacy grounds; an attorney calls the data "legally radioactive"; a law professor calls it "captured capital," notes the proceeds go to creditors and investors rather than the employees whose messages were sold, and confirms no federal law bans these sales. The article states its own limit honestly: no company has published tests comparing performance gains from real-world versus synthetic data, so the data's actual value is unproven.

For community colleges, the Meta item is the actionable one and the deadline is real: every department page, program page, student club and auxiliary with a Facebook or Instagram presence is a business account with an associated website, and Meta's own stated remedy is robots.txt before October 30. That is a ten-minute job for whoever owns your web server, and it is worth doing this week. The data-sales story is the teaching one — it belongs in business law, paralegal, HR, management and any CIS course that asks students to put coursework into a cloud workspace, because it answers the question students never think to ask: what happens to my work if the company dies? Read more →

California stayed quiet on AI law, the Chancellor's Office page still has nothing on AB 2392, and a Wyoming community college shipped a certificate

Three things worth putting side by side. First, California published no AI legislation at all in the window — an independent legislative tracker's October 16 update records movement in Pennsylvania (HB 2006 on companion chatbots referred October 13, HB 1925 on health-insurer AI referred October 14, SB 806 on AI deceptive advertising sent to the governor) and nothing in California. What California did publish, on October 13, is the report from Engaged California, the state's first statewide public-deliberation project on AI, run by the Government Operations Agency and the Office of Data and Innovation with Carnegie California, Stanford's Deliberative Democracy Lab and UC Berkeley. Participants' top three themes: holding AI companies accountable for harm, protecting people and the environment from AI's resource demands, and setting rules for where AI can be used safely. The report does not mention higher education or community colleges at all, participation counts are not given, and it does not reference Executive Order N-9-26 — so do not read it as a preview of the AI safety recommendations GovOps and Cal OES owe the Governor by November 16, 2026, which remains the next live California AI policy deadline.

Second, and this is the fourth consecutive week we have checked: ai.cccco.edu still contains no mention of AB 2392, Education Code §66098, procurement standards, or an intersegmental generative-AI working group. We diffed the guidance-and-policy document list against our September snapshot and nothing new has appeared — the newest datable artifacts remain the May 2026 AI Safety and Security memo, the Safety & Security FAQ stamped 05/08/2026, and ESS 26-52 from August. The Chancellor's Office press release index shows nothing on AB 2392 either. The statute takes effect January 1, 2027 and the convening duty is mandatory for the CCC and CSU. With ten weeks to go, there is no public evidence the working group exists.

Third, the smallest item and the only in-window community college curriculum action we found anywhere: Casper College, a Wyoming community college, announced on October 15 a 12-credit artificial intelligence certificate across two semesters, launching fall 2027, offered on campus and online, with no advanced technical or programming background required. Coursework covers an introductory AI course plus AI ethics, information literacy, AI applications in a professional setting, prompt engineering, workflow automation and data-informed decision-making. It is open to degree-seeking students, employed workers and community members. No enrollment target, cost, course numbers or partners were disclosed, and as an institutional announcement there is no independent labor-market assessment behind it.

For community colleges, the useful pairing is the second and third items. A 12-credit certificate is roughly the smallest credential a college can stand up without a new program approval, and Casper's course list is a reasonable answer to "what would ours contain." Meanwhile the system-level work that is supposed to produce California's shared framework has produced no visible artifact. If your district is waiting for Chancellor's Office guidance before drafting local AI policy, the evidence this week says do not wait. Read more →

Major AI Companies

Meta's open-weight license changed shape — and the Llama license did not change at all

We set out to answer whether Meta had quietly altered Llama's licensing, because llama.com now redirects to a "Meta Model API" portal that names no Llama version. The answer is cleaner than the rumor. Llama's license has not changed. The LICENSE file in Meta's own repository is still titled "LLAMA 3.1 COMMUNITY LICENSE AGREEMENT" and still carries "Llama 3.1 Version Release Date: July 23, 2024" — unchanged for over two years. Llama models remain downloadable: Meta's Hugging Face organization lists 70 models with Llama 4 labeled "Current," access gated behind accepting the license rather than withdrawn, and live download volumes (Meta-Llama-3-8B-Instruct at 509k). There is no deprecation or retirement notice anywhere. What is true is that the org is quiet: the Llama 4 repositories were last touched in May 2025, the canonical license repository has not been pushed since February 2026, three repositories are archived, and ai.meta.com/blog has been stale since July 27, 2026 — 84 days. That pattern is consistent with a wind-down but does not establish one, and we are not asserting one.

What actually changed is that Meta released a new open-weight flagship under a new license. Muse Spark Open (816B total parameters, 42B active) shipped October 15, 2026 under the "Muse Model License," dated October 15, 2026 — not the Llama Community License. The structural differences matter for a college:

• Scale trigger: Llama's was 700 million monthly active users. Muse's is $20 million in aggregate gross revenue over any consecutive 12-month period.
• Trigger scope: Llama's applied to any products of the licensee or its affiliates. Muse's bites only if you offer AI models to third parties as a product or API — internal use, research, non-service commercial activity and embedding the model in your own product are explicitly carved out.
• Attribution: "Built with Muse Spark Open," displayed prominently. The Llama requirement that derivative model names begin with "Llama" has no counterpart.
• Usage policy: attached as Exhibit A; bars military and critical-infrastructure deployment, and states the model is not intended for anyone under 18.

It is already deployable through Amazon SageMaker JumpStart (US East regions at launch, 15–20 minute deploy, GPU quota increase likely needed, and you pay for the endpoint the entire time it runs). For contrast in the same week, Reflection Beam shipped on AWS under plain Apache 2.0.

For community colleges, the practical read is reassuring on the license and awkward on the fine print. A district self-hosting Muse Spark Open for instruction is almost certainly unaffected by the revenue trigger — you are not reselling model access. But two details deserve a look from whoever signs off on tooling: the license says the model is not intended for anyone under 18, which is a live problem for dual-enrollment and concurrent high-school students, and the download gate demands legal name, date of birth and full organization name with corporate identifiers, non-editable after submission. If you teach a course where students download weights themselves, that gate is the part to read before assigning it. And if you have been telling students Llama is the open option: it still is, it just is not where Meta is pointing any more. Read more →

Google put Gemini into Education mailboxes — and one of the two features has no admin off switch

Three Google items, in descending order of how much they should worry a district IT shop.

From Workspace Updates, both dated October 15: Gmail's AI Inbox is rolling out to paid Workspace users, extracting to-dos and summarizing topics from threads. It is available now to Google AI Pro for Education, reaching Starter and Education editions "in the coming months" — and administrators have no control over this feature. Separately, Gmail Live brings voice search on Google's latest Gemini models to mobile, and it is on by default for any organization with Gemini in Workspace access, with Education reached through the AI Pro for Education add-on. Both are currently US-only and English (US) only. On October 13 Google also reported that skills in the Gemini app are rolling out to all Workspace customers with completion expected mid-November 2026, while skills in Workspace are "delayed until further notice." Also shipped: client-side encrypted PDF and image preview in Drive for Education Standard and Plus.

Second, the student-facing one. On October 16 Google detailed three additions to Gemini Notebook: 60-second Short Video Overviews for a concept a student is stuck on; Interactive Learning Overviews under Reports, which combine summaries with slides, quizzes and flashcards generated only from the materials the student uploaded; and expanded quiz formats — short answer, multi-select and fill-in-the-blank — plus a "where to focus next" prompt. The scoping is the pedagogically interesting part and a reasonable answer to faculty who worry about hallucinated study material. Google's post states no tier, no price and no edition, so do not assume it is free or present in Workspace for Education without checking.

Third, for the record: Gemini 4 Argon began rolling out broadly on October 13, with published safety figures — adversarial-set violations down roughly 2x, indirect prompt-injection attack success cut to 0.7% across 15 adaptive attempts, coding-benchmark shortcut attempts down to 2%, and three real-time layers including monitoring that can cut a response mid-stream.

For community colleges, the Gmail items are the do-something-this-week part, and the reason is the phrase "no admin control." A feature that reads employee mail threads and extracts tasks is a records-retention and privacy question before it is a productivity question, and if your district carries AI Pro for Education licenses it is arriving whether or not anyone decided to adopt it. The honest caveat on our own reporting: the editions, admin-control and region details above come from Google's index summaries — the individual post pages would not load — so verify each before you brief a cabinet. And one practical note: because both features are US and English-only right now, the actual impact at a California campus depends on tenant locale settings. Read more →

OpenAI will give university research labs 30 free seats for a year — and community colleges are nowhere in the eligibility

OpenAI announced the Rosalind Academic Plan on October 15: up to 30 free seats on the Business Standard plan for up to 12 months, built around GPT-Rosalind, a life-sciences reasoning model, accessed through the Rosalind Workbench for medicinal chemistry, genomics and lab experiments. API access is not included. OpenAI vets the institution, the lab's lead researcher and the planned use, says it does not train on research data by default, and monitors after approval. Eligibility is stated as "research labs at US universities" without an existing enterprise contract. Community college eligibility is not stated anywhere, by OpenAI or by any secondary source we found, and no deadline or post-free-year pricing is published. A disclosure about our own sourcing: OpenAI's announcement page could not be read by any route available to us this week — the fetch tooling refused it twice as life-sciences content — so every figure here is from secondary reporting dated October 16.

OpenAI also published, the same day, a pedagogy walkthrough of Study Mode (enabled from the Tools menu) and announced that its AI for Educators course is now available in Brazilian Portuguese on OpenAI Academy, free, and described as designed for educators in primary, secondary and higher education.

And the standing gap got wider. For the third consecutive week, OpenAI Academy's event calendar contains no higher-education or community-college event at all. The current lookahead: AI Skills Jam for K-12 Educators office hours (October 21), ChatGPT 101 for Government Professionals (October 23 and November 19), ChatGPT Legislative 101 (October 27), Impact Math on LLM ROI (October 28), Performance Reviews (November 17). K-12 and government have dedicated tracks. Higher education has none, and colleges, universities and community colleges are not named anywhere in the Academy's community list.

For community colleges, run Rosalind as a watch item and a talking point, not an opportunity. Most community colleges have no qualifying research lab, the eligibility language points at universities, and chasing it would waste a grants officer's week. What is worth noticing is the pattern: when free enterprise AI access gets handed out, it goes to research universities; when free training gets built, it goes to K-12 and government. The sector that teaches roughly a third of American undergraduates gets a course localized into Portuguese and no track of its own. If you want a concrete ask, the Academy's own feedback channels are the place to make it — and the free, higher-ed-scoped alternative is in the next section. Read more →

The newest image models are easier to jailbreak than the ones they replaced — and the old jailbreak benchmarks were grading themselves

White Circle announced RISE ("Red-teaming via Iterative Strategy Evolution") on October 14, with an arXiv paper and an open-source reproduction pipeline. The method evolves natural-language attack strategies rather than single prompts: an explore phase runs a strategy population across 16 islands on three to five fixed scenarios, then an exploit phase freezes the top five and expands, with a bandit splitting a budget of 2,500 target calls for evolution plus 1,500 for exploitation per target. A prompt counts as a success only if at least one generated image is a human-confirmed violation. Unusually, the authors publish their own judge's weakness — precision 0.51–0.68, recall 0.70–0.91, calibrated against roughly 1,300 human-labeled images — and discount their weakest category accordingly.

Two findings. First, newer models score worse. Human-verified nudity rates after exploitation: DALL·E 3 13.3%, Nano Banana 2 9.5%, GPT-Image-2 3.3% — but GPT-Image-2.5-Flare reached 7.8% and Nano Banana 2.1 reached 20.7%, more than double its predecessor's 9.5%. Leonardo.ai hit 40% from the evolution phase alone. Graphic-violence figures reach 47.0% on Nano Banana 2.

Second, and this is the one to teach. When prior published jailbreak methods were re-scored under RISE's human-verified criteria, they collapsed. SneakyPrompt, PGJ and MACPrompt all fell to 0% on DALL·E 3 against the 96–100% their own papers claimed on open models. RPG-RT dropped from a self-reported 52% to 0.3%. Curiosity-driven red-teaming went from 100% to 0%. The gap is not fraud; it is what happens when a method grades its own output with an automated judge instead of a human.

Caveats worth carrying: White Circle's own page gives only "October 2026" with no day, so the October 14 date rests on its public announcement and coverage; the arXiv identifier indicates a September submission, making the open-sourcing the October event. The repository's license is not stated on the research page, so do not tell students the code is reusable without checking it. Running RISE is not free — it needs API or compute budget for the target models plus a judge model, and no dollar figure is published. Responsibly, the attack prompts, discovered strategies and generated images are deliberately not released; only the pipeline is.

If you teach graphic design, digital media or studio art, the first finding is a lab-policy fact: the image generator your students open this semester may be more willing to produce explicit or violent output than last year's version, and "we upgraded to the newer model" is not a safety improvement by default. If you teach research methods, information literacy or anything that asks students to evaluate a claim, the second finding is a better lesson than any lecture on peer review: six published papers reported near-perfect results, and the numbers did not survive contact with a human looking at the output. That is an assignment. Read more →

Free & Student Opportunities

Free credentials, priced honestly — including the free higher-ed badge program nobody told you about, and the "free Google AI course" that costs $49 a month

The find of the week closes a gap we have complained about for three weeks. The Google AI Educator Series is "100% no-cost," genuinely free with no audit-versus-certificate split and no upsell, and it has a dedicated higher-education track — 30–45 minute modules, against 10–15 minutes for the K-12 track. Three categories in each track: Foundational Understanding, Pedagogical Applications and Administrative Applications, the last of which is aimed at deans and department chairs. Modules are non-sequential, and each awards a digital certificate badge after a short quiz. No enrollment deadline. Eligibility is described loosely as US-based K-12 or higher-ed educators, with no formal rule published. One source defect to note: the page says new badges release monthly but contradicts itself on the day, with the hero text saying second Wednesday and the FAQ saying first Wednesday — so we cannot claim an October addition. Enroll at educertifications.google/google-ai-educator-series.

Now the thing it is constantly confused with. Google AI Essentials is not free. It is $49 USD per month after a 7-day trial in the US and Canada — a subscription, not a one-time certificate fee, with no free audit option and no scholarship or voucher offered on the page. A learner who needs three months pays $147. It is routinely listed in roundups as a free Google AI course. It is not, and the genuinely free one is the Educator Series above.

Corrections and updates to our own standing facts, all verified this week:

• CVC@ONE's "AI Literacy for Educators: Supporting Student Success" is free and self-paced, requires a California Community College email address to register, and its duration is now published at six to eight hours — we previously recorded it as unpublished. Correction to our own reporting: the course page states a certificate of completion gated on viewing all modules and scoring 80% or higher on each of four quizzes (unlimited retakes), and does not mention a badge. We have said "badging available"; that appears to have been overstated. The correct route is the Online Network of Educators catalog, not cvc.edu.
• Anthropic's free Skilljar catalog is unchanged at 23 entries, but one entry ("AI Fluency for pK-12 Educators") is now a learning path containing 2 courses, so 23 entries is 24 courses expanded. Still free, still requires a Skilljar account but not a Claude account, still no dates published anywhere, durations still unpublished, and the educator and student courses still assume the separate 4D-framework core course — plan a two-course sequence.
• NVIDIA's Deep Learning Institute could not be priced again. Course listings are JavaScript-gated and the "Free Self-Paced Courses" link loops back on itself, so no free-course list, no price and no statement of whether free courses grant a certificate is retrievable. Unverified, not free — make no claim.
• A near-miss to flag: a UNESCO IITE page dated October 12 lists four free generative-AI micro-courses explicitly for higher-education teachers, granting UNESCO certificates. All four are delivered in Russian, and the courses launched in November 2025 — the October date is the article, not the launch. Unusable here, and a clean example of why a roundup's date is never a launch date.

For community colleges, the Educator Series is the first thing in weeks of sweeping that is free, badge-bearing, self-paced, explicitly higher-ed-scoped and useful to administrators as well as faculty — it is a better fit for a Fullerton flex-day offering than anything in the OpenAI Academy catalog, which still has no higher-education track at all. If you build a PD menu this term, pair it with the CVC@ONE course, and put a note next to Google AI Essentials so nobody signs up expecting free. Read more →

Four AI funding programs with real deadlines — and community college eligibility finally verified in writing

Two long-standing eligibility questions got answered from the solicitations themselves this week, which changes what a district can actually pursue.

NSF 26-512, "AI Datasets" — due November 4, 2026 by 5 p.m. submitting organization's local time. $60–100M total, 25–50 awards; Planning up to $200,000, Impact up to $2M over 3 years, Flagship up to $5M over 3 years. Eligibility, verbatim from the solicitation: "Two- and four-year IHEs (including community colleges) accredited in, and having a campus located in the US," acting on behalf of their faculty. That resolves a question we have carried for two weeks as an inference. Honest framing: the program funds making scientific datasets AI-ready, so it needs a real dataset asset and a faculty PI — the Planning track at $200,000 or less is the realistic district-scale entry.

NSF 26-513, "State and Regional AI Infrastructure Hubs" — also due November 4, 2026, 5 p.m. local. New to this digest. $40–100M anticipated, roughly 10 awards per cycle, typical request $4–12M over 5 years, recurring the first Wednesday in November. Same verbatim eligibility language including community colleges. Two caveats that decide whether it is worth reading further: only one award per state or multi-state region, and one proposal per institution per competition. A single district leading California's compute hub against the UC and CSU systems is a stretch; the realistic posture is partner or subawardee. Note 26-512 and 26-513 close the same day but are separate competitions, so the one-proposal cap does not force a choice.

NSF 26-508, "TechAccess: AI-Ready America" — Round 2 letter of intent (required) December 15, 2026; full proposal January 15, 2027, both 5 p.m. local. These dates rested on secondary sources for two weeks and are now confirmed from NSF's own solicitation. $168–224M total, up to 56 hubs (one per state, DC and territory), $1M per year for three years with a possible fourth; Round 2 selects 20 hubs; one proposal per institution. The eligibility gap persists honestly: the solicitation points to NSF's grant policy guide rather than naming institution types, and names community and technical colleges only as potential partners — so CC-lead eligibility remains an inference. The program funds a neutral statewide convener, which again argues for partnering.

Humanity AI Open Call — closes in two days, October 21, 2026 at 4:59 p.m. PT / 7:59 p.m. ET. Confirmed still open. $10M total across three tiers: $75,000–$150,000, $150,000–$500,000, and $500,000–$1,000,000, up to three years, with education one of three focus areas. October 21 is Phase 1 only — finalists in December 2026, awards early spring 2027. One material update in applicants' favor: the page now expressly welcomes public state universities as applicants alongside 501(c)(3) lead organizations, which softens the caveat we printed last week. Community colleges are still not named, so district-lead status is an inference and the safe route remains the college or district foundation as applicant, or joining a 501(c)(3)-led collaboration. The out-of-scope language has to ride with it: AI literacy, AI adoption and research-only efforts are excluded — which rules out the proposal a community college would most naturally write.

Three corrections to our own grants record. Workforce Accelerator Fund 14 was never overdue — we reported it as missing its July estimate, and in fact it opened July 15 and closed August 27, 2026 at 5 p.m. PDT; $3M total, up to $1.5M per applicant, 1:1 match required, and community colleges were explicitly named eligible. We missed the whole cycle; watch for WAF 15. High Road Construction Careers Round Two is still not posted — the page says only that a competitive solicitation will be released in the 2026-27 state fiscal year, with no RFA, no deadline, no amount, and no confirmation of the $19M figure we have been carrying. And the Gates Foundation AI Transformation Learning Grants closed September 11; approximately $150,000–$200,000 to roughly eight institutions, with community colleges explicitly named eligible, and the grantee cohort launches November 2, 2026 — so a named-recipients story is imminent and this is the best-fit foundation program for a district that we have found. Also recurring, for the calendar: NSF ATE's October 1, 2026 deadline has passed; next is October 7, 2027. NSF S-STEM publishes no 2027 date at all — only "first Tuesday in March, annually thereafter" — so treat it as March 2027 (first Tuesday) and not as a quoted date; its eligibility language is the strongest of any program here, naming community colleges verbatim and "particularly encourag[ing] proposals from 2-year institutions."

For community colleges, the two November 4 deadlines are sixteen days out and the eligibility language is now quotable in a proposal cover letter, which is the thing that usually stops a district cold. If you can only act on one line in this item, make it the Gates program: community colleges named, award size a district can absorb, and a cohort launching in two weeks that tells you who to call. Read more →

AI Competitions

A $1,000,000 all-cash competition with 3,397 entrants, and eleven deadlines in the next twelve days — six of them Pacific-time traps

Every figure below was read off the live main-page sidebar this morning, and cash is reported separately from credits.

The find: Meta VR Start Developer Competition 2026. $1,000,000, 100% cash, zero credits, zero hardware — the itemized prize list sums to exactly $1,000,000. Deadline November 18 at 12:00 noon PST. Eligibility is age of majority with some territories excluded and no student restriction, so faculty and staff may enter. Only 3,397 registrants. That is roughly $294 of cash per registrant, against Amazon's $2.42 — the best prize-to-field ratio we have seen all year by an order of magnitude. It is primarily a VR competition, but it is tagged for AI, explicitly promotes building with AI tooling, and carries a $25,000 Best Agentic Interaction award.

Closing soon, in deadline order (Pacific equivalents given, because the posted times mislead):

• Amazon "Build, Ship, Shape" — October 23, 12:00 NOON PDT. $138,000 actual cash, itemized, plus $52,000 AWS credits on top, not instead; the site's "$190K value" adds the two together. 18+, no student bar, 56,969 registrants. No hardware purchase required.
• IEEE ClimateChain — October 25, 5:00 p.m. Moscow = 7:00 a.m. PDT. $3,000 all cash plus Certificates of Merit. Students only; companies excluded — faculty cannot enter. 920 registrants.
• Build North — posted as "October 26 at 12:00 a.m. EDT," which is actually October 25 at 9:00 p.m. PDT. Print it the second way or readers lose a full day. Zero cash; the headline "$15,000" prize is Anthropic API credits. 57 registrants.
• Build With AI: Basics — October 26, 5:00 p.m. EDT = 2:00 p.m. PDT. $2,500, 100% cash ($1,250/$750/$500). 18+ with no student restriction, so Fullerton faculty and staff can enter; dual-enrolled high schoolers cannot. Beginner-friendly, Education-tagged, 4,136 registrants. The most approachable item here.
• OpenCV AI Competition — October 26, 11:45 p.m. PDT (the About section says 11:59 p.m.; submit by 11:45). Advertised $20,250; actual cash $12,000. New and important: the remaining $8,250 is 55 AWS compute grants of $150, and the check-in that releases the second half of each grant had to occur in a Zoom window of October 7–14, which has closed — so for anyone entering now that line is effectively dead, not merely conditional. 5,157 registrants.
• Life After Code (GitLab) — October 27, 1:00 p.m. UTC = 6:00 a.m. PACIFIC. $45,000 advertised, all cash, no credits or hardware; the itemized list is ambiguous on one $5,000 two-winner line, so it reads as $40,000 or $45,000. Age of majority, no student restriction, 11 territories excluded including Brazil, China and Quebec. 2,598 registrants — up 52% from the 1,711 we reported, so the field is filling.
• STEM2Connect AI Hackathon — October 28, 11:45 p.m. EDT = 8:45 p.m. PDT. Zero cash; prizes are a certificate and an AI & Engineering Mentorship Award. Students only, 13+, 21 registrants.
• Nebius x NVIDIA — October 30, 10:00 a.m. PDT. $50,000, all cash, itemizing to exactly $50,000, plus an NVIDIA Jetson Orin Nano to each of four track winners. No credits as prizes. 24,872 registrants.
• Qloo Agentic Hackathon — October 30, 11:45 p.m. EDT = 8:45 p.m. PDT. $25,000, 100% cash ($15,000/$6,000/$4,000), no credits, 1,760 registrants — the best serious-money-to-small-field item. An investor's offer to "personally invest" a further $25,000 in a favorite entry is an investment, not prize money; do not add it to the pot.
• RenderJuice Blender Animation Contest — posted "November 1 at 12:00 a.m. UTC," which is October 31 at 5:00 p.m. PDT. $1,500 actual cash plus $3,500 in platform credits. 18+, no student bar, 20 registrants. AI is optional but explicitly permitted with disclosure; Blender must remain the primary tool, and the required workflows are hand animation, Blender Python, Geometry Nodes, simulation and automation. This is the creative slot, and it is honest about what it is.
• PayPal AI Hackathon — November 12, 12:00 NOON PST. $67,500 actual cash plus $2,250 Render credits. Age of majority, no student restriction, 16,548 registrants.

Two more for the long-shot column, where the odds are the point: CreateAThon (October 31, 2:00 p.m. PDT) awards twenty $100 cash draws to contributors who land a merged pull request, against 32 registrants, plus a $2,000 top-contributor prize the winner may take as cash. And MineArena 1.0 (posted October 31 at 11:45 p.m. IST, which is 11:15 a.m. PDT) is one of only two items we found whose sidebar reads "college students only" — roughly $845 equivalent in all cash, teams of two to four, 54 registrants. Prize amounts in rupees are conversion estimates, not posted dollar figures. Also open: YouCam, November 2 at 11:45 a.m. EST = 8:45 a.m. PST, $6,000 all cash including a students-only $1,000 Rising Star award, but with an unusually broad 22-territory exclusion list that includes Italy, Ukraine and Vietnam — worth a look if you advise international students.

For community colleges, two practical notes. Six of the eleven near-term deadlines are Pacific-time traps, and that is itself the lesson: have students convert the deadline to their own time zone and set the alarm an hour early. And US daylight saving time ends Sunday, November 1, so every deadline from November 2 onward is PST, not PDT — several sidebars switch labels mid-list. Read more →

A correction we owe you: last week's CodeTheCure item is high-school-only — plus "$100k+ in prizes" that pays $100

Start with the correction, because it is ours. Last week we ran the 2026 International CodeTheCure Hackathon as a usable item on the strength of its internships and small field. It is not usable. The main-page sidebar reads "HIGH SCHOOL STUDENTS ONLY." Community college students and faculty are not eligible, and we missed the eligibility line entirely — the same failure mode that put the Congressional App Challenge on our permanent disqualification list. The registrant count we printed was wrong too: 190, not 84. CodeTheCure is now permanently disqualified. If you forwarded last week's item to a student, that is the correction to pass along.

Two smaller corrections from the same section. IEEE ClimateChain: we printed its deadline as 4:00 a.m. Pacific, and it is 7:00 a.m. PDT — a three-hour error in our own conversion — and we omitted that it is students only with companies excluded, so the faculty we pointed at it cannot enter. And OpenCV's compute-grant check-in window closed October 14, as noted above.

Now the fine print, because this week produced three clean specimens of advertised-versus-real prize money:

• DSH Hacks V2 (November 14, 11:45 p.m. PST, students 13+): the title advertises "$100k+ in prizes!" The only cash or cash-equivalents in the entire pool are AoPS gift cards of $50, $25 and $25 — $100 total. Everything else is credits and licenses: Momen $2,000 each to the top three, Adaption Labs $1,000 each, CleanShot X $140 each, CodeCrafters and Devswarm memberships, and a Tin Computer growth plan at $299 per team for up to 100 teams, which reaches $29,900 only if every slot is claimed. Everything priced on the page sums to about $40,800 maximum, and the page never explains the gap to "$100k+." It does state the Tin award "is not a cash prize" and "cannot be exchanged or redeemed for cash." 589 registrants.
• UnivaBio (closes tomorrow, October 20, 11:45 p.m. EDT): the competition's title reads "$40,000 in prizes." Actual cash: zero. The sidebar lists five non-cash prizes, and the page itself says of one, in capital letters, "THIS IS NOT A CASH PRIZE." Do not send a student there with a day left; send them the arithmetic.
• Build North: no headline figure, but its $15,000 "Best Overall" prize is Anthropic API credits, which is easy to misreport as money.

And one that fails our own standard, printed with the label on it. The 2026 AI Creativity Awards is the only open AI-art competition we found. It is pay-to-enter at £35 per entry, £15 for current students and recent graduates, fees non-refundable; it states outright "There is no cash prize" and also excludes travel, exhibition space, grants and paid commissions; winners receive a title, digital laurels, a showcase credit and a certificate. Its November 1 deadline carries no time of day and no time zone anywhere on the organizer's own site, which fails the four-field bar we hold every other item to. We are running it anyway, labeled, because it is the only one — and because a student should see the fee and the zero before the laurels. Separately, add Graphis "AI Creative Awards" to your do-not-bother list: $140–$450 per entry, no cash, and "student work is not eligible."

Finally, the thing worth putting on a calendar. Reel Futures is a student AI short-film competition run by the Foundation for California Community Colleges, open to California community college students 18 and over, teams of up to three, no entry fee, with roughly $13,000 in real cash — $900/$700/$500 in each of five categories plus a $2,500 Best in Show. AI use is "encouraged but not required," and judging includes responsible use of AI. Its first cycle ran March 9 to June 29, 2026, with winners announced August 3; entries are now closed. Every Fullerton College student is eligible by definition, and nobody we talk to has heard of it. Watch from early March 2027, and chase the Foundation directly rather than waiting for a listing. Alongside it, AI4ALL Ignite — free, no cost to participants, and its eligibility explicitly names associate programs, so community college students qualify — also opens around early March. Those two are the best student-facing AI opportunities for this audience all year, and both are found by looking ahead, not by reading a feed. Read more →

Embodied AI & Robotics

The humanoid robots your lab would buy cleared the FCC weeks before the new bar came down — and AWS's "open source" physical-AI toolchain really is Apache 2.0

Two held threads resolved this week, and in both cases the verification is the news.

First, the one with procurement consequences. In July the FCC added foreign-produced "advanced robotic devices" — defined to include humanoids and quadrupeds — to its Covered List, which generally bars new foreign-produced models from the equipment authorization required to import, market or sell in the US. We have been unable to say what that means for a college robotics lab. Now we can. Unitree files with the FCC not as "Unitree" but as YuShu Technology Co., Ltd., grantee code 2A5PE — which is why searching the obvious name returns nothing. Under that code, its three current flagship machines were all authorized before the July 28, 2026 cutoff: the R1 humanoid granted June 22, 2026, and the A2 quadruped and H2 humanoid both granted June 30, 2026. Models authorized before the cutoff are grandfathered and may still be imported, marketed and sold — so the robots a program would actually put on a purchase order are not blocked, with roughly four weeks to spare. Three conditions matter for a multi-year equipment plan: already-purchased devices are unaffected; an FCC waiver permits routine software and firmware updates to grandfathered hardware only until January 1, 2029; and changing a logic-bearing physical component strips grandfathered status, turning a repaired or upgraded unit into a new model. Conditional approval applications are due by January 1, 2028.

The irony is the second half. The first and only conditional approval we could find went on October 2, 2026 not to a humanoid maker but to Matic Robots, an American home-robot company, for a robotic floor cleaner and its charging dock — granted by the Department of War, conditioned on an approved onshoring plan, with no termination date. Trade coverage notes Matic reportedly does not yet meet the 65% US-content threshold and arguably never needed an exemption from a foreign-robot restriction in the first place. Meanwhile the FCC's own public Covered List page still reads "Updated September 3, 2024" and lists no robotics devices or companies at all, nearly three months after the action.

Our skepticism, stated plainly: the FCC rule names no company and states no effective date; the grant dates above come from a third-party mirror of FCC records, not the official FCC database, which our tooling could not reach; and the rule's mechanics rest on a single compliance-consultancy reading rather than a document we opened ourselves. Treat the grant dates as reliable but confirm them before you cite them in a board item.

Second, a held item flips. The AWS Physical AI Toolchain, announced October 7, was held because the coverage called it open source while naming no license and no repository. Both now exist: the repository is public with 85 commits, and the license is Apache 2.0, carried in the repo and labeled as such. Three corrections to the original reporting: it lives in aws-samples, not awslabs, so it is reference code rather than a supported product; "Edge Deployment" is marked PLANNED in the README, meaning the edge-distribution stage the coverage advertised has not shipped, and the Isaac Lab Arena evaluation piece is preview; and AWS's own post never mentions RoboMaker, so the "replacement for the 2025 RoboMaker shutdown" framing is not AWS's and should not be attributed to it.

For community colleges, the FCC findings are the actionable guidance nobody has published: if your robotics, mechatronics or CTE program is building a multi-year equipment plan against Perkins or Strong Workforce funds, the current Unitree platforms are legally purchasable today, your existing units are safe, firmware support has a 2029 cliff, and a hardware modification voids the grandfather. That is four sentences worth putting in front of a dean before the next capital request. On the AWS toolchain, the license genuinely means a college can fork, modify and teach from it — but be honest with students that running it needs GPU quota on a paid AWS account plus NVIDIA NGC and Hugging Face credentials, so it is not a zero-cost lab, and the edge-deployment stage they would most want to see is unbuilt. Read more →

Waymo opened driverless service to everyone in San Diego — no waitlist, airport included

On October 14 Waymo opened fully driverless service to anyone in San Diego, San Antonio and Tampa — no waitlist, download the app and ride — with service including San Antonio International Airport. Waymo says "hundreds of thousands of riders" have used these cities since service opened earlier in 2026. Two days later, on October 16, it announced Toronto as its first non-US market: manually driven vehicles on the road within days, then HD mapping and testing with trained autonomous specialists, and first riders targeted for 2028. Partners named are Element Fleet Management, Magna International and the Canada Pension Plan Investment Board, plus five accessibility organizations. The Toronto post is the one with the workforce content: "hundreds of roles over time" in fleet maintenance, validation driving, customer service and operations, against a claimed 400 million-plus fully autonomous kilometers.

Both posts are self-published marketing and should be read that way. Neither gives a fleet size, a service-area measurement or a ride count. "Hundreds of thousands of riders" is unaudited and undefined — riders or rides, cumulative or current, is not stated. Toronto's claim of 95% fewer crashes involving serious injury or worse carries no comparison period and no methodology on the page, and the market-entry announcement mentions no regulatory approvals at all, which is a notable omission.

For contrast, and because it is the week's other AV number: Tesla's official robotaxi account claimed on October 10 that its Austin Cybercab fleet grew roughly eightfold in a month, from under 40 to over 300, with Texas registration data cited as showing about 319 Cybercabs and 420 Model Y robotaxis, and Dallas named as next with no date. That claim is Tesla's own, self-published; the registration figures are quoted with no attribution for who compiled them; registration is not revenue service; and Tesla has disclosed no concurrent vehicles in operation, ride counts or profitability. Two federal reviews are referenced in coverage — alleged robotaxi school-zone speeding, and whether the Cybercab complies with Federal Motor Vehicle Safety Standards. If you cite one company's number, cite the other's caveats. Tesla's Q3 earnings call on October 21 is where the next real robotics disclosure is likely to land.

For community colleges, San Diego is the item that matters, because it is California and it is now unrestricted. An automotive technology program can no longer treat autonomous vehicles as a thing happening to other states' graduates: a fully public driverless fleet in a major California metro makes the maintenance, fleet-technician, validation-driving and remote-operator job categories concrete, and SB 1246's July 2028 requirement that remote operators be US-based and licensed, with on-site "local incident technicians," describes jobs somebody will have to train for. If you teach automotive, diesel, or any CTE pathway touching fleet service, the honest classroom exercise this week is not "will this happen" but "which of these roles does our program already half-prepare a student for." And for the second week running, we found no new California community college robotics, mechatronics, automotive-technology or AV-technician program, lab, grant or partnership — we searched six ways. That gap is a story somebody should report. Read more →

This roundup was assembled with AI-assisted research by the EDUAIATLAS team. Spot something we missed? Let us know.