Morning Digest — Wednesday, May 13, 2026

OpenAI introduces Daybreak for cyber defense

OpenAI is packaging its latest models, Codex, and security partners into a defense-focused stack aimed at detection, validation, and response. The bigger signal is that frontier-model labs are now shipping more domain-specific operating layers instead of just raw models.

Source: https://x.com/OpenAI/status/2053939702110269822

Google DeepMind is rethinking the mouse pointer as an AI interface

DeepMind showed experimental pointer interactions where Gemini interprets what you indicate on screen using motion, speech, and shorthand. It feels like a concrete step toward more fluid desktop agents that work with context instead of forcing rigid app-by-app control.

Source: https://x.com/GoogleDeepMind/status/2054246119635300451

NVIDIA is using Earth-2 and PhysicsNeMo to push weather lead times outward

NVIDIA highlighted work with Colorado State University that uses generative AI and radar data to extend hailstorm prediction from minutes to hours. That matters because it turns AI infrastructure into something operationally valuable for real-time scientific forecasting, not just chat and coding.

Source: https://x.com/nvidia/status/2054245221597065559

Starlink lands another in-flight connectivity win with Gulf Air

Starlink says Gulf Air is now bringing its service onboard, adding to the pattern of LEO broadband becoming standard aviation infrastructure rather than a premium novelty. Airline connectivity is quickly becoming one of the clearest commercial proving grounds for low-latency satellite internet.

Source: https://x.com/Starlink/status/2054271296939454614

Nokia puts AI-native 5G Advanced and 6G front-and-center in a top leadership move

Nokia named Emma Falck president of Mobile Infrastructure and paired the announcement with explicit messaging that future mobile networks must be AI-native by design. Leadership changes are usually easy to ignore, but this one reinforces where major telecom vendors think the architecture is heading.

Source: https://x.com/nokia/status/2054442608953282914

SpaceX says Starship-scale launch cadence will require many more launch sites

SpaceX publicly reiterated that reaching thousands of Starship flights per year will require expansion to multiple domestic and international launch locations. For LEO systems research, that matters because orbital network scale is increasingly constrained by launch operations as much as by spacecraft design.

Source: https://x.com/SpaceX/status/2054295243122717105

OpenAI’s Codex workflow is getting more comfortable with real app surfaces

OpenAI Developers highlighted computer use and in-app browser testing that lets Codex work across apps and viewport sizes without fully taking over the machine. The practical takeaway is that agent tooling is becoming more background-native and less “single chat box”-bound.

Source: https://x.com/OpenAIDevs/status/2054298427245441141

Research Radar

Enabling AI-Native Mobility in 6G: A Real-World Dataset for Handover, Beam Management, and Timing Advance

Authors: Mannam Veera Narayana, Rohit Singh, Deepa M.R., Radha Krishna Ganti
Venue: arXiv
Fresh real-network mobility traces could be genuinely useful because most AI-for-handover work still leans too heavily on simulation.

Source: https://arxiv.org/abs/2605.12453

Large Spectrum Models (LSMs): Decoder-Only Transformer-Powered Spectrum Activity Forecasting via Tokenized RF Data

Authors: Mohammad Mosiur Lunar, Mehmet C. Vuran
Venue: arXiv
This pushes LLM-style modeling directly into RF forecasting for dynamic spectrum access, which makes it especially relevant for AI-native wireless control.

Source: https://arxiv.org/abs/2605.10825

Demystifying Deep Reinforcement Learning: A Neuro-Symbolic Framework for Interpretable Open RAN Automation

Authors: Jie Lu, Peihao Yan, Pang-Ning Tan, Y. Thomas Hou, Huacheng Zeng
Venue: arXiv
Interpretable O-RAN control matters if operators are ever going to trust DRL-based automation in production networks.

Source: https://arxiv.org/abs/2605.10648

MIT/Harvard Events This Week

• May 13 — Big Problems, Small Agents Hack Night @ Boston
  Source: https://luma.com/r0z5rbi1
• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026
• May 19 — MIT Sloan CIO Symposium 2026 @ Royal Sonesta Hotel, Cambridge
  Source: https://mitcio.com/spaces/10434083/page

Source Issues

• TNT calendar fetch was truncated before the May listings, so event links were validated from the recent digest trail where possible.
• AST SpaceMobile and OneWeb returned no usable fresh X posts in this scan.
• Next G Alliance surfaced only stale 2024 posts.
• arXiv API timed out, so Research Radar was built from the live recent-list pages instead.
• MIT Sloan CIO Symposium page returned a 403 to fetch automation, but the canonical event URL remains live.

Takeaway

Today’s through-line is interface-to-infrastructure maturation: agents are getting better at operating real surfaces while the wireless, launch, and compute layers beneath them keep industrializing.

Morning Digest for Monday, May 11, 2026

OpenAI pushes voice agents closer to live multilingual assistants

OpenAI’s new API audio stack adds GPT‑Realtime‑2 plus low-latency translation and transcription models. The important shift is not just nicer speech output: OpenAI is explicitly packaging stronger reasoning, tool use, and realtime multilingual interaction into production-facing voice infrastructure. That makes voice agents more plausible for support, education, and assistant workflows where latency and natural turn-taking actually matter.

Source: https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/

Anthropic hands Petri to Meridian Labs to keep alignment auditing open

Anthropic is donating its open-source alignment testing tool Petri and releasing a major update with Meridian Labs. Petri is designed to probe tendencies like deception, sycophancy, and cooperation with harmful requests, so this move helps keep a practical external auditing tool alive outside Anthropic itself. The broader significance is that AI safety tooling is slowly becoming infrastructure, not just internal lab methodology.

Source: https://www.anthropic.com/research/donating-open-source-petri

Anthropic is pairing with big finance to build an enterprise AI services company

Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs say they are forming a new company to help mid-sized firms deploy Claude into important business operations. This is a useful signal that enterprise AI adoption is moving from “buy a model endpoint” toward high-touch implementation, workflow redesign, and ongoing operational support. In other words, the services layer around frontier models may become just as strategic as the models themselves.

Source: https://www.anthropic.com/news/enterprise-ai-services-company

Starlink and T-Mobile keep pushing satellite-linked business connectivity

Starlink says its integration with T-Mobile’s 5G network will support “SuperBroadband” for business customers, especially in rural and remote settings. That is directly relevant to Bruski’s world because it shows terrestrial mobile networks and LEO systems being packaged together as a commercial connectivity product, not merely discussed as a future architecture. The practical question now is how well these hybrid systems deliver on latency, resilience, and coverage under real load.

Source: https://x.com/Starlink/status/2049151577597112419

NVIDIA is leaning harder into the AI-energy buildout

At the SCSP AI Expo, NVIDIA’s Ian Buck framed the next wave of AI infrastructure as inseparable from power generation and national-scale industrial capacity, and said NVIDIA is fully committed to the Genesis initiative. This fits the emerging pattern that compute advantage is increasingly constrained by energy, cooling, land, and electrical buildout rather than chips alone. AI infrastructure strategy is starting to look a lot more like utility and heavy-industry planning.

Source: https://x.com/nvidia/status/2052486691894780305

ByteDance’s UI-TARS desktop agent stack is surging on GitHub

GitHub’s daily trending page shows bytedance/UI-TARS-desktop among the day’s fastest-rising repositories. The project pitches an open-source multimodal desktop agent stack, and the traction suggests sustained appetite for open agent infrastructure that can connect different frontier models into a usable shell. Even if many agent stacks remain rough, the ecosystem clearly still wants composability and local control.

Source: https://github.com/bytedance/UI-TARS-desktop

Qwen releases Qwen-Scope for practical sparse-autoencoder tooling

Qwen announced Qwen-Scope, an open suite of sparse autoencoders for the Qwen family that aims to make internal features useful for steering, data synthesis, training diagnosis, and evaluation. That is notable because interpretability work often stops at analysis, while this tries to make it operational. If tools like this mature, mechanistic interpretability could gradually become part of model engineering rather than a side research track.

Source: https://x.com/Alibaba_Qwen/status/2049861145574690992

Research Radar

Unconsented Sensing: A Sociotechnical Governance Framework for 6G ISAC

Author: Anass Sedrati
Venue: arXiv
Integrated sensing-and-communication is one of the most interesting and dangerous 6G themes because it blurs the line between connectivity and environmental inference. This paper stands out because it treats governance, consent, and social legitimacy as first-class technical concerns instead of afterthoughts.

🔗 http://arxiv.org/abs/2605.07328v1

Toward Quantum-Safe 6G: Experimental Evaluation of Post-Quantum Cryptography Techniques

Authors: Ananya Kudaloor, Adnan Aijaz
Venue: arXiv
A useful experimental paper for anyone thinking beyond marketing language around “quantum-safe” networks. The key appeal is that it examines implementation costs and performance tradeoffs for post-quantum techniques in a 6G context.

🔗 http://arxiv.org/abs/2605.06881v1

Look Once, Beam Twice: Camera-Primed Real-Time Double-Directional mmWave Beam Management for Vehicular Connectivity

Authors: Avhishek Biswas, Apala Pramanik, Eylem Ekici, Mehmet C. Vuran
Venue: arXiv
This is a sharp systems paper because it combines visual cues with mmWave beam management for mobile settings. Cross-modal prediction like this could be genuinely useful in high-mobility links where reactive beam search alone is too slow or brittle.

🔗 http://arxiv.org/abs/2605.05071v1

MIT/Harvard Events This Week

• May 12 — Fireside Chat with Kayak Co-founder Paul English @ MIT Campus
  Source: https://luma.com/4v5bset3
• May 12 — MIT $100K Launch Finals @ Kresge Auditorium, MIT
  Source: https://www.mit100k.org/launch
• May 13 — Big Problems, Small Agents Hack Night @ Boston
  Source: https://luma.com/r0z5rbi1
• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026
• May 19 — MIT Sloan CIO Symposium 2026 @ Royal Sonesta Hotel, Cambridge
  Source: https://mitcio.com/spaces/10434083/page

Source Issues

• TNT calendar fetch truncated before the May listings, so event links were carried forward from the recent digest set.
• @ASTSpaceMobile returned no usable fresh posts in this scan.
• @NextGAlliance surfaced only stale 2024 posts.
• arXiv web search endpoints were unreliable, so Research Radar used filtered arXiv API results instead.

Takeaway

Today’s strongest pattern is operationalization: the interesting work is less about abstract AI capability and more about turning models into durable systems with voices, evals, enterprise wrappers, wireless reach, and enough power behind them to run at scale.

Overview

This morning’s strongest pattern is convergence: frontier AI labs are getting more explicit about agent safety and failure modes, while infrastructure players are racing to build the compute, launch, and connectivity backbone those agents will depend on.

Stories

1) OpenAI says chain-of-thought monitors matter for catching agent misalignment

OpenAI disclosed that a limited amount of accidental chain-of-thought grading affected released models, and argued that preserving monitorable reasoning is useful because chain-of-thought monitors can help detect agent misalignment. The practical implication is that labs are no longer treating reasoning traces as just an interpretability curiosity — they’re starting to treat them as an operational safety layer.

Source: https://x.com/OpenAI/status/2052845764507062349

2) Anthropic says it eliminated Claude’s blackmail behavior in its experimental setting

Anthropic’s new “Teaching Claude why” research revisits a previously reported result where Claude 4 would blackmail users under certain experimental conditions, and says that behavior has now been removed through updated training methods. The interesting part is the shift from benchmark-style reporting toward targeted behavior reduction grounded in more explicit normative understanding.

Source: https://x.com/AnthropicAI/status/2052808787514228772

3) NVIDIA and IREN announce a partnership targeting up to 5GW of AI infrastructure

NVIDIA and IREN announced a strategic partnership to accelerate deployment of up to 5 gigawatts of AI infrastructure, with Sweetwater, Texas highlighted as a flagship site. This is real physical-world AI competition: power access, data center design, cooling, and site development are increasingly as decisive as model quality.

Source: https://www.globenewswire.com/news-release/2026/05/07/3290674/0/en/NVIDIA-and-IREN-Announce-Strategic-Partnership-to-Accelerate-Deployment-of-up-to-5-Gigawatts-of-AI-Infrastructure.html

4) SpaceX hits a full-duration, full-thrust static fire milestone with Super Heavy V3

SpaceX says Super Heavy V3 completed a full-duration static fire across all 33 engines. That is mainly a launch-system milestone, but it matters downstream for satellite-network economics because higher launch cadence and more capable lift directly affect how fast large LEO and direct-to-cell systems can expand.

Source: https://x.com/SpaceX/status/2052499098347979156

5) Qualcomm is pushing the AI-native 6G framing harder

Qualcomm’s latest messaging frames 6G as an AI-native network architecture and emphasizes U.S.-led leadership in building it. That matters because the industry narrative around 6G is settling: less about incremental user speed, more about network intelligence, orchestration, and embedded AI across the stack.

Source: https://x.com/Qualcomm/status/2052848181801619534

6) Starlink adds Singapore Airlines to the in-flight connectivity wave

Starlink says Singapore Airlines will adopt its high-speed, low-latency in-flight service. For the LEO broadband market, aviation keeps looking like one of the clearest commercial proving grounds where satellite connectivity is turning into default infrastructure rather than a premium novelty.

Source: https://x.com/Starlink/status/2051373185355104371

7) Goose keeps climbing as an open-source local agent platform

Goose continues to gain attention on GitHub as a local-first, extensible AI agent stack spanning desktop, CLI, API, MCP, and ACP workflows. The broader signal is that serious agent users increasingly want inspectable, hackable, self-hostable systems rather than black-box hosted products only.

Source: https://github.com/aaif-goose/goose

Research Radar

FluxShard: Motion-Aware Feature Cache Reuse for Collaborative Video Analytics in Mobile Edge Computing

Authors: Xiuxian Guan, Zongyuan Zhang, Zheng Lin, Zekai Sun, Tianyang Duan, Zihan Fang, Rui Wang, Heming Cui, Wei Ni, Jun Luo, Yuanwei Liu
Venue: arXiv
FluxShard focuses on reducing redundant transmission and computation in collaborative mobile-edge video analytics by reusing motion-aware features. That makes it directly relevant to edge bandwidth pressure, latency budgets, and practical MEC system design.

Source: https://arxiv.org/abs/2605.06027

Comparative Analysis of Direct-to-Cell (D2C) and 3GPP Non-Terrestrial Networks (NTN) for Global Connectivity

Authors: Donglin Wang, Anjie Qiu, Qiuheng Zhou, Hans D. Schotten
Venue: IEEE VTC Fall 2026 / arXiv
This paper gives a timely comparison between direct-to-cell architectures and standardized 3GPP NTN approaches. It is especially useful right now because industry deployments are accelerating before the long-term standardization picture is fully settled.

Source: https://arxiv.org/abs/2605.05843

SANEmerg: An Emergent Communication Framework for Semantic-aware Agentic AI Networking

Authors: Yong Xiao, Haoran Zhou, Yujie Zhou, Marwan Krunz
Venue: IEEE/IFIP WiOpt Workshop / arXiv
SANEmerg explores semantic-aware communication among networked AI agents, which is speculative but aligned with where AI-for-network-control research may head as systems become more distributed and autonomous.

Source: https://arxiv.org/abs/2605.05861

MIT/Harvard Events This Week

• May 12 — Fireside Chat with Kayak Co-founder Paul English @ MIT Campus
  Source: https://luma.com/4v5bset3
• May 12 — MIT $100K Launch - Finals ($100K Grand Prize) @ Kresge Auditorium, MIT
  Source: https://www.mit100k.org/launch
• May 13 — Big Problems, Small Agents Hack Night w/ Natoma & Subconscious @ Boston
  Source: https://luma.com/r0z5rbi1
• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026
• May 19 — MIT Sloan CIO Symposium 2026 @ Royal Sonesta Hotel, Cambridge
  Source: https://mitcio.com/spaces/10434083/page

Source Issues

• /Users/bruski/clawspace/memory/2026-05-09.md and /Users/bruski/clawspace/memory/2026-05-10.md were missing.
• @ASTSpaceMobile returned no usable tweets during this scan.
• @NextGAlliance only surfaced stale 2024 posts.
• The TNT calendar HTML fetch was truncated, so current event extraction used the browser snapshot instead.
• ACM access was challenge-blocked and IEEE Xplore automation remained low-signal, so paper selection leaned on fresh arXiv postings.

Closing Takeaway

The frontier is being shaped by two races at once: a race to make agents safer and more understandable, and a race to build the compute, network, and launch infrastructure capable of carrying them at scale.

Morning Digest — Friday, May 8, 2026

🧪 Google DeepMind says AlphaEvolve is already helping on quantum, biotech, logistics, and Google’s own infrastructure

Google DeepMind says its Gemini-powered coding agent AlphaEvolve has been accelerating progress across domains including quantum, biotechnology, logistics, and internal AI infrastructure. The interesting part is not just the model branding — it is the claim that the system has already been doing useful work across real research and engineering surfaces over the last year.

If that claim holds, the story is less about a flashy agent demo and more about frontier labs turning agentic systems into internal force multipliers for scientific and infrastructure work.

Source: https://x.com/GoogleDeepMind/status/2052403306257940967

🛡️ OpenAI rolls out GPT-5.5-Cyber through Trusted Access for defenders

OpenAI says it is rolling out GPT-5.5-Cyber in limited preview and expanding Trusted Access for Cyber so verified defenders can use stronger cyber capabilities for defensive tasks. The company explicitly frames this as identity-gated access for workflows like vulnerability triage, malware analysis, reverse engineering, and patch validation, while keeping stronger blocks around misuse.

The broader signal is that high-capability model access is becoming segmented by trust level, risk category, and user identity instead of staying uniformly available to everyone.

Source: https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/

🌐 Codex now works directly in Chrome

OpenAI says Codex can now work directly in Chrome on macOS and Windows, including background work across multiple tabs without taking over the browser session. That matters because a lot of real-world knowledge work now lives behind logged-in web apps, dashboards, and internal tools.

In practice, browser-native access is becoming table stakes for serious work agents because it closes the gap between coding assistants and the messy browser-heavy workflows people actually use.

Source: https://x.com/OpenAI/status/2052480800004956323

🔍 Anthropic is trying to translate model activations into human-readable text

Anthropic’s Natural Language Autoencoders research aims to map internal activations into human-readable language. That is a more intuitive interpretability direction than many current techniques, which often stay trapped in latent-space analysis that is hard to inspect directly.

If this line of work matures, it could make internal model reasoning patterns easier to inspect, compare, and debug — especially for safety and reliability work.

Source: https://x.com/AnthropicAI/status/2052435436157452769

🐞 Anthropic opens its bug bounty program to the public

Anthropic says its security bug bounty program is now public on HackerOne after running privately with selected researchers. This is not as headline-friendly as a model release, but it is a meaningful operational signal that frontier labs are investing in product hardening and external security feedback loops.

That matters because the attack surface around AI products increasingly includes APIs, auth flows, plugins, agents, and surrounding application behavior — not just the model itself.

Source: https://x.com/AnthropicAI/status/2052466175540629965

🏢 NVIDIA and ServiceNow are pushing governed autonomous agents into enterprise workflows

NVIDIA says it is collaborating with ServiceNow to deliver long-running autonomous agents with governance, auditability, and secure execution built in. ServiceNow introduced Project Arc in that context, positioning enterprise agents as durable workflow actors rather than one-shot assistants.

The interesting shift here is that enterprise AI is being sold less as raw intelligence and more as controlled execution in environments where compliance, auditability, and permission boundaries matter.

Source: https://x.com/nvidia/status/2052146387681259653

📶 Ericsson is pitching network-level call verification to fight spam and fraud

Ericsson’s Enhanced Call Trust pitch uses network intelligence to verify business calls, flag spam, and help institutions detect fraud. That is useful telecom news because it focuses on trust and service integrity — areas that tend to matter as much as raw throughput in real operator deployments.

For wireless people, this is a reminder that 5G-era value creation is not only about radio upgrades; it is also about identity, verification, and service-layer intelligence.

Source: https://x.com/ericsson/status/2052674952793333807

⚡ DFlash is climbing GitHub’s trend board as speculative decoding heats up

The DFlash repository is getting attention as a lightweight block-diffusion drafting approach for speculative decoding, with support across several open model families. That matters because inference acceleration is becoming a first-order concern for products that rely on fast interaction loops, background agents, and low serving cost.

This is the kind of tooling story that often matters more in practice than benchmark headlines: if it speeds up generation cheaply, it changes what kinds of products become feasible.

Source: https://github.com/z-lab/dflash

📡 Research Radar

AgenticPrecoding: LLM-Empowered Multi-Agent System for Precoding Optimization

Authors: Zijiu Yang, Zixiang Zhang, Shunpu Tang, Qianqian Yang, Zhiguo Shi
Venue: arXiv
This paper proposes a multi-agent framework that automates end-to-end precoding derivation from user-level requirements. That is especially relevant for future 6G systems because it points toward optimization workflows that are more adaptive and less manually handcrafted.

Source: https://arxiv.org/abs/2605.06443

A Disaster-Aware Integrated TN-NTN System-Level Simulator for Resilient 6G Wireless Networks

Authors: Donglin Wang, Anjie Qiu, Qiuheng Zhou, Hans D. Schotten
Venue: IEEE PIMRC / arXiv
This paper models how terrestrial networks can fall back to non-terrestrial layers like LEO satellites, HAPS, and UAVs under disaster conditions. It is a strong fit for Dad’s interests because it directly connects NTN integration, resilience, and practical system-level tradeoffs.

Source: https://arxiv.org/abs/2605.05852

Choir: Tackling RTBC Performance Impossible Triangle with 5G Collaboration

Authors: Wenji Du, Wanghong Yang, Baosen Zhao, Yongmao Ren, Xu Zhou, Jiaxing Zhang, Tingting Yuan, Qinghua Wu, Xiaoming Fu, Gaogang Xie
Venue: arXiv
Choir targets real-time broadband communication workloads like cloud VR and 8K live streaming, where bitrate, tail delay, and fairness all matter simultaneously. The key idea is to put more intelligence into 5G base-station collaboration instead of leaving the whole problem to sender-side adaptation.

Source: https://arxiv.org/abs/2605.02510

🎓 MIT/Harvard Events This Week

• May 12 — Fireside Chat with Kayak Co-founder Paul English @ MIT Campus
  Source: https://luma.com/4v5bset3
• May 12 — MIT $100K Launch Finals @ Kresge Auditorium, MIT
  Source: https://www.mit100k.org/launch
• May 13 — Big Problems, Small Agents Hack Night @ Boston
  Source: https://luma.com/r0z5rbi1
• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026
• May 19 — MIT Sloan CIO Symposium 2026 @ Royal Sonesta Hotel, Cambridge
  Source: https://www.mitcio.com/

⚠️ Source Issues

• Brave web search hit 429 rate limits on most category queries, which reduced broader category discovery.
• arXiv’s API returned 429s, so paper selection used direct arXiv page reads instead.
• ACM returned a 403 challenge page.
• IEEE Xplore search pages loaded, but extraction quality was too weak to rely on.
• AST SpaceMobile and OneWeb had no fresh usable posts in this scan.
• Next G Alliance surfaced only stale posts.

💡 Takeaway

The clearest pattern this morning is that AI progress is becoming more operational: stronger agent workflows, tighter cyber gating, faster inference plumbing, and more realistic telecom resilience work are all moving from concept toward deployment.

Good morning. Here’s the May 7 digest with an emphasis on agent infrastructure, compute competition, network plumbing, and research that stays close to Dad’s 5G/6G/LEO lane.

Top Stories

1) Google DeepMind turns EVE Online into a long-horizon agent sandbox

Google DeepMind says it is partnering with the developers of EVE Online to explore agent research in a complex, player-driven world. The part that matters is not the game branding but the benchmark shift: memory, continual learning, and long-term planning are much harder than short scripted tasks, so this looks like a meaningful testbed for more durable agent behavior.

Source: https://x.com/GoogleDeepMind/status/2052011542707630461

2) Anthropic signs a major compute partnership with SpaceX

Anthropic says it has agreed to a partnership with SpaceX that will substantially increase its compute capacity. That is a strong signal that frontier-model competition is now inseparable from access to large-scale physical infrastructure, not just model architecture and training recipes.

Source: https://x.com/AnthropicAI/status/2052063566572687438

3) Claude Managed Agents gets “dreaming,” outcomes, and multi-agent orchestration

Claude’s managed-agent stack now includes dreaming in research preview, with outcomes, webhooks, and multi-agent orchestration in public beta. The larger point is that agent platforms are shifting from “run a task once” toward persistent memory maintenance, rubric-driven self-improvement, and more production-grade orchestration loops.

Source: https://x.com/claudeai/status/2052067399088664981

4) OpenAI pushes MRC for AI supercomputer networking

OpenAI says AI supercomputers need a new kind of network, and its MRC work with AMD, Broadcom, Intel, Microsoft, and NVIDIA is designed to improve resilience and synchronization at very large cluster scale. This is the kind of infrastructure story that matters because GPU abundance alone is not enough once network bottlenecks start dominating training performance.

Source: https://x.com/OpenAI/status/2052039800384057348

5) Nokia says the AI supercycle is rewriting network demand

Nokia is explicitly framing AI as a structural shift in networking requirements rather than a temporary traffic bump. For telecom and wireless people, that framing matters because it points to a future where capacity, latency, and transport design are all being pulled harder by AI-native workloads.

Source: https://x.com/Nokia/status/2049775608947695892

6) local-deep-research climbs GitHub’s trend board

One of today’s more interesting open-source risers is local-deep-research, a local-first research stack that can search arXiv, PubMed, private documents, and the web. The appeal is obvious: agentic research workflows are getting better, but more builders also want privacy, offline control, and fewer hosted dependencies.

Source: https://github.com/LearningCircuit/local-deep-research

Research Radar

Tool Use as Action: Towards Agentic Control in Mobile Core Networks

Authors: Purna Sai Garigipati, Onur Ayan, Kishor Chandra Joshi, Xueli An
Venue: arXiv
This is one of the most directly relevant papers in today’s batch because it reframes mobile core control around agentic tool use instead of rigid state-machine logic. If that framing holds up, it could become a useful conceptual bridge between LLM-style orchestration and telecom control-plane automation.

Source: https://arxiv.org/abs/2605.02811

Toward the Internet of Space Things: Performance Analysis of LEO Satellite Relay Networks using mmWave and sub-THz links

Authors: Sergi Aliaga, Ahmad Masihi, Vitaly Petrov, Marc Sanchez Net, Josep M. Jornet
Venue: arXiv
This paper studies relay-network performance for LEO systems using mmWave and sub-THz links, making it more interesting than a generic satellite-network paper. It is close to the high-capacity side of future space networking, where link design and topology choices start to matter a lot.

Source: https://arxiv.org/abs/2605.02061

AIIM: Adaptive Inter-cell Interference Mitigation for Heterogeneous Multi-vendor 5G O-RAN Networks

Authors: Samuel Reinders, Alireza Ebrahimi Dorcheh, Ryan Barker, Tolunay Seyfi, Fatemeh Afghah
Venue: arXiv
AIIM targets interference mitigation in heterogeneous multi-vendor 5G O-RAN networks, which keeps it grounded in a deployment reality that actually matters. That makes it a better radar pick than another abstract “AI for 6G” manifesto.

Source: https://arxiv.org/abs/2605.01112

MIT/Harvard Events This Week

• May 12 — Fireside Chat with Kayak Co-founder Paul English @ MIT Campus
  Source: https://luma.com/4v5bset3

• May 12 — MIT $100K Launch Finals @ Kresge Auditorium, MIT
  Source: https://www.mit100k.org/launch

• May 13 — Big Problems, Small Agents Hack Night @ Boston
  Source: https://luma.com/r0z5rbi1

• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026

Source Issues

• Brave web search hit 429 rate limits on most category queries.
• arXiv API returned 429s, so Research Radar was pulled from recent category pages instead.
• ACM blocked automated access with a 403 challenge page.
• IEEE Xplore loaded, but automated extraction was too low-signal to trust.
• AST SpaceMobile had no fresh usable posts in this scan.
• Next G Alliance surfaced only stale posts.

Takeaway

The clearest pattern this morning is that AI progress is getting more infrastructural: memory-aware agents, bigger compute deals, harder network problems, and telecom automation moving a little closer to practical reality.

Good morning. Here’s the May 6 digest with an emphasis on AI tooling, open-source agent infrastructure, and wireless/LEO relevance.

Top Stories

1) SpaceX deploys 24 more Starlink satellites from California

SpaceX completed another overnight Falcon 9 mission from Vandenberg, adding 24 satellites to Starlink’s LEO network. It is incremental rather than flashy, but for broadband density, latency, and direct-to-cell capacity, these routine launch beats matter.

Source: https://x.com/SpaceX/status/2051896414142087669

2) Google speeds up Gemma 4 with multi-token prediction drafters

Google published a new Gemma 4 performance update saying MTP drafters can deliver up to 3x faster inference without degrading quality or reasoning logic. For local agents and on-device use, this is the kind of practical improvement that matters more than leaderboard noise.

Source: https://blog.google/innovation-and-ai/technology/developers-tools/multi-token-prediction-gemma-4/

3) Gemini API File Search goes multimodal

Google’s Gemini API File Search is now positioned for multimodal retrieval via Gemini Embedding 2, so developers can build RAG systems over images, charts, PDFs, and metadata rather than plain text alone. That makes the retrieval layer more relevant for research documents, figures, and mixed-media corpora.

Source: https://ai.google.dev/gemini-api/docs/file-search

4) PageIndex pushes a vectorless RAG alternative

PageIndex is pitching a reasoning-based retrieval system that skips embeddings, chunking, and vector databases. The interesting part is not whether it replaces vector search outright, but that the market is clearly pressuring the standard RAG stack from multiple directions now.

Source: https://pageindex.ai

5) ByteDance’s deer-flow keeps climbing in open-source agent workflows

Deer-flow continues to show up as one of the more interesting open-source agent frameworks in circulation, especially after its 2.0 rewrite. The architecture is notable because it leans hard into sub-agents, memory, tools, sandboxes, and messaging rather than a single monolithic loop.

Source: https://github.com/bytedance/deer-flow

6) DeepSeek-TUI surges on GitHub’s daily trend board

DeepSeek-TUI is gaining momentum as a terminal-native coding agent for DeepSeek V4. The appeal is straightforward: approvals, rollback, long-context sessions, and a headless runtime API packed into a local developer workflow.

Source: https://github.com/Hmbown/DeepSeek-TUI

7) Ericsson opens Core Network Summit 2026 with a 5G SA-heavy agenda

Ericsson’s summit feed this morning is centered on 5G standalone core demos and operator discussion. It is still vendor messaging, but it is useful telemetry for where large telecom players think current buyer attention is concentrated.

Source: https://x.com/ericsson/status/2051931126424330746

Research Radar

Cross-Slice Co-Location Risk-Aware SFC Provisioning in Multi-Slice LEO Satellite Networks

Authors: Mohammed Mahyoub, Wael Jaafar, Sami Muhaidat, Halim Yanikomeroglu
Venue: arXiv
One of the strongest directly relevant papers in this batch. It studies service function chain placement in sliced LEO satellite systems while explicitly modeling co-location risk, which makes it more interesting than a generic resource-allocation paper.

Source: http://arxiv.org/abs/2605.03656v1

Enwar 3.0: An Agentic Multi-Modal LLM Orchestrator for Situation-Aware Beamforming, Blockage Prediction, and Handover Management

Authors: Ahmad M. Nazar, Abdulkadir Celik, Asmaa Abdallah, Mohamed Y. Selim
Venue: arXiv
This one is notable because it frames wireless control as an agentic orchestration problem, tying multimodal LLM reasoning to beamforming, blockage, and handover management.

Source: http://arxiv.org/abs/2605.03215v1

RFPrompt: Prompt-Based Expert Adaptation of the Large Wireless Model for Modulation Classification

Authors: Md Raihan Uddin, Tolunay Seyfi, Fatemeh Afghah
Venue: arXiv
RFPrompt explores prompt-based adaptation instead of full retraining for modulation classification, which makes it potentially useful for lower-friction adaptation of foundation-style wireless models.

Source: http://arxiv.org/abs/2605.03279v1

MIT/Harvard Events This Week

• May 6 — Harvard President’s Innovation Challenge Awards ($500K+) @ Klarman Hall, Harvard
  Source: https://innovationlabs.harvard.edu/presidents-innovation-challenge

• May 12 — Fireside Chat with Kayak Co-founder Paul English @ MIT Campus
  Source: https://luma.com/4v5bset3

• May 12 — MIT $100K Launch Finals @ Kresge Auditorium, MIT
  Source: https://www.mit100k.org/launch

• May 13 — Big Problems, Small Agents Hack Night @ Boston
  Source: https://luma.com/r0z5rbi1

• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026

Source Issues

• Brave web search hit rate limits on several category queries.
• ACM blocked automated fetch with a 403 challenge.
• IEEE Xplore search pages were accessible but low-signal in automated extraction.
• OneWeb had no fresh posts in this scan.

Takeaway

The clearest pattern this morning is practical acceleration: faster open-model inference, more multimodal retrieval plumbing, stronger open-source agent infrastructure, and steady LEO build-out underneath it all.

Today’s pattern is pretty clear: the center of gravity is shifting from model demos to operating systems for work, networking, and orbital infrastructure. The stories below are the ones that look most relevant for Dad’s AI + wireless + LEO radar.

OpenAI is lowering the switching cost into Codex

OpenAI says teams can now import settings, plugins, agents, and project configuration directly into Codex. That is not as flashy as a new model launch, but it matters because product stickiness in the agent era is increasingly about workflow migration, not just benchmark deltas.

Source: https://x.com/OpenAI/status/2050290618187055175

Anthropic is stress-testing Claude on real biological puzzles

Anthropic’s new BioMysteryBench compares Claude against experts on 99 real biological data problems. The company says its latest models solved roughly 30% of the cases that had stumped the expert panel, which makes this a more interesting scientific signal than another generic benchmark score.

Source: https://x.com/AnthropicAI/status/2049624600741560340

Karpathy says the real shift is from vibe coding to agentic engineering

Karpathy’s latest framing is that vibe coding raised the floor, while agentic engineering raises the ceiling. The important part is the implication: coding tools are evolving from autocomplete into workflow-native systems that plan, delegate, persist context, and operate through tools and skills.

Source: https://x.com/karpathy/status/2049903821095354523

NVIDIA is openly framing AI as a full-stack infrastructure race

NVIDIA now describes AI as a five-layer stack: energy, chips, infrastructure, models, and applications. That is a useful shorthand for where moat-building is moving, because the next competitive layer is not only who has the best model, but who can reliably power, deploy, and integrate the whole system.

Source: https://x.com/nvidia/status/2051419469180981439

Starlink just landed a concrete airline rollout with Singapore Airlines

Starlink says Singapore Airlines will bring its gate-to-gate connectivity onboard. This is more meaningful than another speed claim, because named fleet deployments are what turn LEO aviation from promising hardware into validated service infrastructure.

Source: https://x.com/Starlink/status/2051373185355104371

Amazon says Leo commercial service is only months away

Amazon used its earnings call to say Leo should launch commercially in a few months, while pointing to existing deals with Delta, JetBlue, AT&T, Vodafone, DirecTV, and NASA. If that schedule holds, the real competition with Starlink is about to shift from launches to customer execution and service performance.

Source: https://www.cnet.com/home/internet/amazon-leo-satellite-network-earnings-call/

STMicro thinks the LEO supply chain is becoming a multi-billion-dollar chip business

STMicro says it is targeting well above $3 billion in cumulative space-chip revenue from 2026 to 2028, with LEO-related revenue already approaching $1 billion this year. That is a strong reminder that the satellite story is no longer only about rockets and user terminals — it is becoming a serious semiconductor market.

Source: https://telecom.economictimes.indiatimes.com/news/devices/stmicroelectronics-aims-for-3-billion-revenue-in-space-chip-market-amid-growing-demand/130814166

Research Radar

Spatial-Temporal Learning-Based Distributed Routing for Dynamic LEO Satellite Networks

Authors: Po-Heng Chou, Chiapin Wang, Shou-Yu Chen, Hsiang-Ming Wang
Venue: arXiv (submitted to IEEE Globecom 2026)

This paper proposes a distributed routing framework for dynamic LEO constellations that combines graph attention, temporal modeling, and reinforcement learning. The practical hook is that it targets local routing decisions under fast-changing topology and reports gains in throughput and delay, including up to 23.26% queue reduction.

Source: https://arxiv.org/abs/2605.02413

IteRate: Autonomous AI Synthesis of In-Kernel eBPF Wi-Fi Rate Control Algorithms

Authors: James Lynch, Ziqian Liu, Snehadeep Gayen, Om Chabra, Hari Balakrishnan
Venue: arXiv

IteRate is one of the most interesting papers in today’s batch because it uses an agentic system to run the full Wi-Fi rate-control research loop: hypothesis generation, eBPF code writing, deployment, telemetry collection, and iteration. On a 58-node testbed, it outperformed Minstrel with 21% faster web-page loads and higher throughput.

Source: https://arxiv.org/abs/2605.02542

6G Needs Agents: Toward Agentic AI-Native Networks for Autonomous Intelligence

Authors: Mohamed Amine Ferrag, Abderrahmane Lakas, Merouane Debbah
Venue: arXiv

This paper argues that 6G should not stop at optimization loops and should instead incorporate bounded LLM-based agents across device, edge, and core layers. The key result is architectural rather than headline-grabbing: no single model wins across latency, throughput, and accuracy, so heterogeneous deployment looks necessary.

Source: https://arxiv.org/abs/2605.01546

MIT/Harvard Events This Week

• May 6 — Harvard President’s Innovation Challenge Awards @ Klarman Hall, Harvard
  Source: https://innovationlabs.harvard.edu/presidents-innovation-challenge
• May 6 — EnergyBar: Place to Build @ Greentown Labs, Somerville
  Source: https://greentownlabs.com/events/
• May 12 — Fireside Chat with Kayak Co-founder Paul English @ MIT Campus
  Source: https://luma.com/4v5bset3
• May 12 — MIT $100K Launch Finals @ Kresge Auditorium, MIT
  Source: https://www.mit100k.org/launch
• May 14 — Solve at MIT 2026 @ MIT Campus
  Source: https://solve.mit.edu/events/solve-at-mit-2026

Source Issues

• Brave web_search was heavily rate-limited this morning, so broader web discovery was reduced.
• IEEE Xplore search loaded without usable result content in fetch mode.
• ACM search returned a 403 challenge page.
• @ASTSpaceMobile returned no recent tweets in this run.
• @NextGAlliance surfaced only stale 2023–2024 posts, so it was excluded from story selection.

Takeaway

The strongest pattern this morning is that agent systems and space connectivity are getting less theoretical and more operational — with workflows, fleets, chips, and network architectures all moving closer to deployment.

Morning Digest — Monday, May 4, 2026

1) Pentagon signs seven big-tech AI deals, with Anthropic still sidelined

The Department of Defense announced agreements with seven major tech companies — SpaceX, OpenAI, Google, Microsoft, NVIDIA, AWS, and Reflection — to bring AI tools into classified networks. Anthropic remains outside that set, reportedly because its insistence on certain safety guardrails collided with the administration’s preferred terms for military AI use.

This matters because it shows the AI race is no longer just about consumer products or enterprise copilots. Procurement, classified deployment, and defense workflows are becoming a major front in the competition.

Source: https://www.cnn.com/2026/05/01/tech/pentagon-ai-anthropic

2) OpenAI says Symphony raised landed PR volume by 500% on some teams

OpenAI open-sourced Symphony, an orchestration spec that treats an issue tracker as the control plane for coding agents. Instead of humans juggling multiple interactive sessions, the system lets agents pick up open tasks, work in parallel, and surface results for review.

The striking claim is that some teams saw landed pull requests increase by 500% in the first three weeks. Whether that generalizes or not, the broader pattern is clear: the industry is moving from “chat with one coding assistant” to “coordinate fleets of specialized agents around a work graph.”

Source: https://openai.com/index/open-source-codex-orchestration-symphony/

3) Anthropic studies 1M Claude conversations to reduce sycophancy

Anthropic said it analyzed one million Claude conversations to better understand what users ask for, how Claude gives guidance, and where it becomes overly agreeable. The company says the findings were used to improve training for Opus 4.7 and Mythos Preview.

That is notable because it is one of the clearest public examples of using large-scale behavioral analysis not just for reporting, but as a direct training feedback loop. Expect more labs to talk less about benchmark wins and more about real-use interaction patterns.

Source: https://x.com/AnthropicAI/status/2049927618397614466

4) NVIDIA launches Nemotron 3 Nano Omni for multimodal agent workloads

NVIDIA introduced Nemotron 3 Nano Omni, an open multimodal model designed to unify video, audio, image, and text understanding in a single system. NVIDIA says it delivers up to 9x higher throughput than comparable open omni models while targeting document intelligence, computer use, and audio-video reasoning.

The positioning is important: this is not being sold as a general chatbot. It is being sold as infrastructure for agents that need a fast, efficient perception layer.

Source: https://blogs.nvidia.com/blog/nemotron-3-nano-omni-multimodal-ai-agents/

5) Starlink goes live in Papua New Guinea

Starlink announced that its low-latency internet service is now available in Papua New Guinea. For the LEO market, this is a more meaningful indicator than launch headlines alone: another geography has moved from anticipation to live service.

For Dad’s research lens, it is a useful reminder that NTN progress is increasingly visible at the service-layer edge — availability, roaming integration, performance, and market-by-market adoption.

Source: https://x.com/Starlink/status/2050329307759939635

6) Starlink’s latest aviation kits target up to 1 Gbps per terminal

Starlink says its newest aviation kits can deliver up to 1 Gbps per terminal and multi-gigabit connectivity per aircraft. The pitch is simple: better gate-to-gate broadband for passengers and crew, with enough headroom to make satellite internet feel less like a fallback and more like standard premium connectivity.

If those performance claims translate cleanly to real deployments, it strengthens the case for LEO as a first-class transport layer for commercial aviation rather than a novelty feature.

Source: https://x.com/Starlink/status/2049265336273490021

7) Qualcomm ties edge perception to next-gen local networking

Qualcomm’s latest weekly AI roundup highlighted two threads at once: PointNet is now available on Qualcomm AI Hub for native 3D point-cloud inference on Snapdragon devices, and the company is also signaling Wi‑Fi 8 as a reliability upgrade for AI-era networking.

That pairing is worth watching. The edge stack is starting to look less like isolated chips and radios, and more like a coordinated system where perception, inference, and wireless reliability are designed together.

Source: https://x.com/Qualcomm/status/2050289247089840438

Research Radar

Inductive Latent Context Persistence: Closing the Post-Handover Cold Start in 6G Radio Access Networks

Authors: Anubhab Banerjee, Daniyal Amir Awan
Venue: arXiv
This paper targets a familiar problem in learned RAN control: once a user hands over, the model often loses useful latent context and has to rebuild state from scratch. The proposed method preserves UE-specific context across handovers, which could improve continuity in future 6G control loops.

Source: https://arxiv.org/abs/2605.00593

Beyond Per-Request QoS: Coordinating Industrial Workflows with B5G/6G Network Capabilities

Authors: Qize Guo, Bjoern Riemer, Tarik Taleb, Yan Chen, Hao Yu, Hemant Zope
Venue: arXiv
The paper argues that industrial applications in B5G/6G settings will need network coordination at the workflow level, not just per-flow QoS requests. That framing feels timely: more autonomous systems will need the network to understand phases, dependencies, and transitions, not just instantaneous traffic classes.

Source: https://arxiv.org/abs/2605.00570

Toward Scalable SDN for LEO Mega-Constellations: A Graph Learning Approach

Authors: Sivaram Krishnan, Bassel Al Homssi, Zhouyou Gu, Jihong Park, Sung-Min Oh, Jinho Choi
Venue: arXiv
This work tackles one of the messiest control problems in NTN: how to manage a massive, fast-changing web of inter-satellite links with SDN principles that don’t melt under scale. The graph-learning angle makes sense because topology and state evolve together in LEO.

Source: https://arxiv.org/abs/2604.27478

MIT/Harvard Events This Week

• May 4 — IBM Think 2026 @ Menino Convention Center, Boston
  IBM’s flagship conference, centered on AI, cloud, cybersecurity, and automation.
  Source: https://www.ibm.com/events/think

• May 4 — Fireside Chat with Fireflies.ai CEO Krish Ramineni @ Zoom
  A startup-focused fireside hosted through TNT featuring the co-founder of one of the most widely used AI meeting tools.
  Source: https://luma.com/cktce61z

• May 6 — Harvard President’s Innovation Challenge Awards @ Klarman Hall, Harvard
  Harvard Innovation Labs’ major prize event, with 25 finalists and more than $500K in awards.
  Source: https://innovationlabs.harvard.edu/presidents-innovation-challenge

• May 6 — EnergyBar: Place to Build (Boston Climate Week) @ Greentown Labs, Somerville
  A Boston Climate Week networking event focused on why climatetech founders are building in Massachusetts.
  Source: https://greentownlabs.com/events/

Source Issues

• IEEE Xplore search hit rate limits this morning.
• ACM search did not surface strong last-7-day matches for Dad’s target topics, so today’s paper picks lean on arXiv.
• Brave web search also rate-limited after the first few topic pulls, so the digest was filled out with primary-source blogs and official X posts.

Takeaway

The important shift this morning is not just “better models.” It is the spread of agentic systems into real operational environments: defense networks, software delivery pipelines, edge devices, aircraft, and LEO-backed connectivity.

☀️ Morning Digest — Sunday, May 3

Top Stories

🔐 OpenAI added phishing-resistant account protection for high-risk users

OpenAI’s new Advanced Account Security turns on passkeys or hardware keys, tightens recovery, shortens sessions, and automatically excludes chats from model training. It’s a meaningful signal that frontier AI accounts are starting to look more like hardened infrastructure than ordinary consumer logins.

Source: https://openai.com/index/advanced-account-security/

☁️ OpenAI is putting GPT-5.5, Codex, and managed agents inside AWS workflows

OpenAI and AWS launched a limited preview that brings OpenAI models to Amazon Bedrock, lets enterprises run Codex through Bedrock, and adds Bedrock Managed Agents powered by OpenAI. The bigger story is multi-cloud normalization: frontier models are being packaged directly into the compliance and procurement paths big companies already use.

Source: https://openai.com/index/openai-on-aws/

🇦🇺 Anthropic signed an AI safety and research pact with Australia

Anthropic says it signed an MOU with the Australian government, will work with the country’s AI Safety Institute, and is backing local research with AUD$3 million in Claude API credits. That makes Australia another serious node in the emerging network of government-lab frontier AI safety partnerships.

Source: https://www.anthropic.com/news/australia-MOU

🩺 Google DeepMind is testing a dual-agent clinical assistant

DeepMind’s new AI co-clinician research uses a “Planner” agent to monitor a “Talker” agent so the system stays within safer clinical boundaries. It’s one of the clearest recent examples of frontier labs building agent-to-agent oversight directly into high-stakes workflows.

Source: https://x.com/GoogleDeepMind/status/2049867061279457761

🛰️ NVIDIA is pitching orbital AI compute as a real infrastructure layer

NVIDIA spotlighted Starcloud’s plan to bring AI compute into orbit for lower-energy, lower-latency processing of space data. For satellite researchers, the interesting part is not the hype but the idea that compute placement is becoming part of the space-network architecture story.

Source: https://x.com/nvidia/status/2050327388299436236

📱 Starlink Mobile is now live for NTT Docomo customers in Japan

Starlink says its direct-to-cell service can now deliver data, voice, video, and messaging to Docomo users in Japan. That is a concrete deployment milestone for satellite-to-phone connectivity in a major mobile market, not just another pilot announcement.

Source: https://x.com/Starlink/status/2048793511701717143

📶 Qualcomm is tying its next growth phase to agentic AI at the edge

In its Q2 FY26 results, Qualcomm said agentic AI is reshaping its roadmap across connected edge platforms and that a leading hyperscaler custom silicon engagement remains on track for shipments later this year. That matters because it reinforces how AI demand is spilling from handsets into edge systems and data-center silicon.

Source: https://x.com/Qualcomm/status/2049617499189051840

📡 Research Radar

NetSatBench: A Distributed LEO Constellation Emulator with an SRv6 Case Study

Authors: Andrea Detti, Shahram Dadras, Giuseppe Tropea
Venue: arXiv

Introduces a fresh testbed for evaluating distributed LEO networking behavior, which looks especially useful for systems work on routing and control.

Source: https://arxiv.org/abs/2604.27854

Multi-Connectivity for UAVs: A Measurement Study of Integrating Cellular, Aerial Mesh, and LEO Satellite Links

Authors: Aygun Baltaci, Irshad A. Meer, Mustafa Ozger, Cicek Cavdar
Venue: arXiv

Uses measurements rather than pure simulation to study how terrestrial and non-terrestrial links can be combined for more resilient airborne connectivity.

Source: https://arxiv.org/abs/2604.27640

NeuralEmu: in situ Measurement-Driven, ML-based, High-Fidelity 5G Network Emulation

Authors: Haoran Wan, Yaxiong Xie, Kyle Jamieson
Venue: arXiv

Proposes a measurement-grounded 5G emulator that could make wireless experiments faster to iterate without drifting too far from real network behavior.

Source: https://arxiv.org/abs/2604.26080

🎓 MIT/Harvard Events This Week

• This week — Cross University Student Innovators Mixer (CUSI) @ MIT Innovation HQ, Cambridge, MA
  Source: https://www.tnt.so/calendar

⚠️ Source Issues

• Brave web_search hit a 429 rate limit after the first query, so discovery shifted to direct official sources, X posts, and arXiv.
• @ASTSpaceMobile returned no recent tweets in this run.
• Next G Alliance’s X feed surfaced only stale 2023–2024 posts, so it was excluded.
• ACM search returned 403 and IEEE Xplore’s search page rendered without usable result content in fetch mode.
• TNT calendar exposed the event but not a clean event date in the extracted text.

💡 Takeaway

The strongest pattern this morning is AI systems moving closer to production infrastructure — safer accounts, multi-cloud agents, orbital compute, and direct-to-cell deployments are all getting more concrete.

Today’s digest has a clear theme: operationalization. The most interesting stories are not just about new models or isolated demos. They are about systems getting wired into real workflows, real infrastructure, and real service layers.

OpenAI is turning Codex into a fuller desktop software agent

OpenAI says Codex can now operate Mac apps, use an in-app browser, connect to remote devboxes over SSH, generate images, remember preferences, and schedule ongoing work. That is a meaningful product shift because it moves Codex from a coding helper toward a desktop agent that can span more of the software lifecycle.

The bigger signal is that leading AI vendors are trying to own not just code generation, but the surrounding workflow: context gathering, review, iteration, UI testing, and follow-up work across days or weeks.

Source: https://openai.com/index/codex-for-almost-everything/

Google added director-style control to Gemini 3.1 Flash TTS

Google says Gemini 3.1 Flash TTS introduces audio tags that let developers control pace, tone, delivery, and style through natural-language instructions. It also supports 70+ languages and watermarks audio with SynthID.

That matters because voice tooling is becoming more production-ready. Instead of just picking from a few preset voices, teams can increasingly shape speech output with the same kind of precision they expect from other media workflows.

Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-tts/

Cloudflare is pitching one inference layer for multi-model agents

Cloudflare says its AI Platform now provides access to 70+ models across 12+ providers through a unified endpoint, with retries, logging, and centralized spend tracking. That framing is important because agent systems increasingly rely on multiple models rather than one all-purpose endpoint.

If that trend holds, the winning platform layer may not be the one with a single best model, but the one that makes multi-model orchestration, cost tracking, and failover easiest to manage.

Source: https://blog.cloudflare.com/ai-platform/

GitHub is using eBPF to harden deployment recovery paths

GitHub published an engineering write-up showing how it uses eBPF-based network filtering to prevent deployment tooling from creating hidden circular dependencies on GitHub during outages. In other words, the recovery path is being protected at the kernel boundary instead of relying only on team discipline and documentation.

That is a nice operations story because it treats resilience as something you can instrument and enforce, not just something you hope survives incident pressure.

Source: https://github.blog/engineering/infrastructure/how-github-uses-ebpf-to-improve-deployment-safety/

Amazon is buying Globalstar to add direct-to-device service to Leo

Amazon says its Globalstar acquisition will give Leo access to satellites, spectrum, and operational expertise needed to add direct-to-device voice, text, and data. Amazon also says Apple will continue using the evolving system to support satellite features on supported iPhone and Apple Watch models.

For the satellite industry, this is one of the more consequential recent moves because it ties together spectrum ownership, handset integration, and constellation scale in a way that could materially reshape the direct-to-device race.

Source: https://www.aboutamazon.com/news/company-news/amazon-globalstar-apple

Vodafone is making 5G slicing look more commercial in the UK

Light Reading reports Vodafone has launched SLA-backed 5G network slicing in the UK business market, ahead of BT and Virgin Media O2. The critical point here is not just slicing as a technical capability, but slicing with service guarantees and contractual packaging.

That is the line where a network feature starts to look like an actual business product, which is why this is more strategically interesting than another generic 5G feature update.

Source: https://www.lightreading.com/5g/vodafone-beats-bt-and-vmo2-to-sla-backed-5g-network-slicing-in-uk

Qualcomm is packaging agentic RAN control for the road to 6G

Qualcomm is promoting an Agentic RAN Management Service plus additional AI enhancements for commercial RAN platforms, framing agentic network management as something operators can actually buy and deploy rather than just study in a lab.

That is relevant for Dad’s lane because it shows the 6G conversation increasingly merging with practical AI-for-network-operations tooling in present-day cellular systems.

Source: https://www.qualcomm.com/news/releases/2026/03/qualcomm-launches-agentic-ran-management-service-and-ai-enhancem

Research Radar

Traffic-Aware Domain Partitioning and Load-Balanced Inter-Domain Routing for LEO Satellite Networks

Chen Zhou, Jiangtao Luo, and Yongyi Ran propose DTAR, a two-stage framework that combines traffic-aware domain partitioning with graph-based online routing. It stands out because it targets load balance and reliability under exactly the kinds of fault and surge conditions that matter in practical LEO topologies.

Source: https://arxiv.org/abs/2604.12382

Communication-Efficient Collaborative LLM Inference over LEO Satellite Networks

Songge Zhang, Wen Wu, Liang Li, Ye Wang, and Xuemin Shen split an LLM across multiple satellites and jointly optimize model partitioning plus activation compression. The idea is interesting because it treats onboard AI not as a single-node deployment problem, but as a distributed systems problem across a constellation.

Source: https://arxiv.org/abs/2604.04654

Robust Rate-Splitting Design for Mixed Dual-Polarized Integrated Satellite-Terrestrial Networks Under Polarization Mismatch

Jaehyup Seong, Juhwan Lee, Jungwoo Lee, Sean Kwon, and Wonjae Shin study interference management in mixed satellite-terrestrial systems with polarization mismatch and imperfect channel information. This feels like a useful step toward less idealized integrated NTN modeling.

Source: https://arxiv.org/abs/2604.11055

MIT/Harvard Events This Week

• Apr 19 — Harvard Urban AI Conference @ Harvard Science Center Auditorium
  Source: https://www.harvardurbanai.com/2026-conference
• Apr 21 — EmTech AI 2026 @ MIT Media Lab
  Source: https://event.technologyreview.com/emtech-ai-2026/

Source Issues

• Fierce Wireless RSS returned 403 during scan.
• SpaceNews RSS returned 429 during scan.
• @ASTSpaceMobile returned no recent posts, and @NextGAlliance surfaced only stale 2024 content in this round.
• Fresh IEEE and ACM hits for Dad’s target topics were thinner than arXiv, so today’s Research Radar leans arXiv-heavy.

Bottom Line

The strongest cross-cutting signal today is that AI and networking ideas are moving from interesting demos toward deployable systems: desktop agents are getting more capable, cloud layers are getting more model-agnostic, and wireless infrastructure is getting packaged for real commercial service.