TL;DR
4 takeawaysMeta shares six math papers written with Muse Spark; it says five answer previously open research questions
Sources: Meta
Aleph Alpha releases Kolibri, an open-weight English-German model with 78B total and 3B active parameters and up to 1M tokens of context
Sources: Aleph Alpha
Cloudflare launches a Web Search API in AI Gateway with Ceramic.ai, Exa and Linkup, whose crawlers must follow its verified-bot rules
Sources: Cloudflare
Short: Meta opens Muse Gadgets for DIY agent hardware, and OpenAI safety-report lead David Robinson resigns, saying its culture is broken
Sources: TechCrunch
The Details
Story 01 / 03
Meta publishes six math papers written with Muse Spark
- What happened
- Meta shared six mathematics papers from a months-long collaboration in which mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, with no custom research scaffold. Meta says five of the papers present answers to previously open research questions, across probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra. In one, Muse Spark generated a search program in GAP that found a 384-element counterexample to a 2024 group-theory conjecture by M. Kida; the researchers verified it and completed the argument. Another settles a wave-collapse question left open in 2015.
- Why it matters
- It’s a concrete, checkable test of AI as a research partner: mathematicians guided the work, a second group reviewed it, and each paper marks which passages were drafted by researchers and which by AI. Meta also notes that teams outside Meta independently announced solutions to some of the same problems.
- Enterprise impact
- The workflow is the takeaway. Use a frontier model to explore and search, keep experts in charge of the problem and the proof, and add independent review before anything ships.
- Startup impact
- AI-assisted research tools are moving from demos to published, credited results. Clear attribution and verification features could set products in this space apart.
- What to focus on
- How the papers hold up under wider peer review, whether other labs publish similarly documented collaborations, and how journals handle passages marked as AI-drafted.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 02 / 03
Aleph Alpha releases Kolibri, a sovereign open-weight model
- What happened
- On the Day of German Reunification, Aleph Alpha released Kolibri, an English-German mixture-of-experts model with 78B total parameters and 3B active, supporting context of up to 1M tokens. Full weights are on Hugging Face under the Apache 2.0 license. Aleph Alpha says Kolibri is specialized for German, reasoning, math and agentic behavior, and is built for mission-critical work in regulated areas including public administration, industry and aerospace.
- Why it matters
- Sovereign AI is turning into real products. Aleph Alpha pitches full supply-chain transparency, from data ingestion to final evaluations, plus a model small enough to run on-premise without sending internal data to third-party inference services.
- Enterprise impact
- European public-sector and regulated buyers get another open-weight option they can host themselves. Test German-language quality and agentic tasks on your own workloads.
- Startup impact
- With Apache 2.0 weights and only 3B active parameters, Kolibri is a practical base for fine-tuning for German-language or compliance-heavy niches.
- What to focus on
- Independent benchmarks against other open models, adoption by public agencies, and what Aleph Alpha’s technical report shows.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 03 / 03
Cloudflare adds a Web Search API to AI Gateway
- What happened
- Cloudflare launched a Web Search API through AI Gateway, starting with partners Ceramic.ai, Exa and Linkup. Developers can pay for web search with AI Gateway credits, log search calls, and control who can use which providers. Partner crawlers must comply with Cloudflare’s published requirements for “Verified bots,” and search responses must link to where the crawled content came from.
- Why it matters
- Agents need fresh context, and search is becoming standard infrastructure for them. Cloudflare is also trying to set a higher bar for how AI crawlers behave toward site owners.
- Enterprise impact
- Bring web search under the same gateway controls, logs and billing as model calls, rather than letting teams wire up separate search keys.
- Startup impact
- Grounded search is getting easier to plug in. Compare providers on result quality and cost per query, and keep source links visible to users.
- What to focus on
- Which providers join next, pricing at scale, and whether other platforms adopt similar crawler standards.
Discuss this story
Comment with your GitHub account. One thread per story.
Worth a Glance
Meta opens Muse Gadgets for DIY agent hardware
Open-source firmware and a Linux SDK let developers build devices that connect to Muse. Meta also made 5,000 Muse Home Links to give free to Muse subscribers.
TechCrunch(opens in a new tab)OpenAI safety-report lead resigns, saying its culture is broken
In an essay in The Atlantic, David Robinson said he led the safety reports that accompanied OpenAI’s major launches. An OpenAI spokesperson said the company pauses training or holds back models when it needs to slow down.
TechCrunch(opens in a new tab)
One Thing Not to Miss Today
Verification is the new feature
Meta’s math papers mark which passages AI drafted and add outside review, Aleph Alpha sells supply-chain transparency, and Cloudflare requires its search partners to cite their sources. Across very different launches, being able to prove where an answer came from is becoming part of the product.