Tech Pulse Daily: OpenAI o3, Ray Serve LLM, PeopleSoft
Curated by Dillip Chowdary - June 19, 2026 - Morning IST edition
Today's Top Highlights
- OpenAI OpenAI says an o3 Deep Research workflow helped specialists revisit 376 unresolved pediatric rare-disease cases and surface 18 confirmed diagnoses.
- Ray Google Cloud and Anyscale published Ray Serve LLM optimizations for GKE with up to 5x higher throughput and 8x lower latency.
- Cloud Google Cloud made Cloud Network Insights generally available for hybrid and multi-cloud network path visibility with Broadcom AppNeta.
- Google Google Cloud published its Europe digital sovereignty position, emphasizing compliance, open collaboration, and customer control.
- OpenAI OpenAI and Molecule.one reported a near-autonomous chemistry workflow where GPT-5.4 improved a difficult medicinal chemistry reaction.
OpenAI o3 Deep Research Reanalyzes Rare Disease Cases
OpenAI says an o3 Deep Research workflow helped specialists revisit 376 unresolved pediatric rare-disease cases and surface 18 confirmed diagnoses.
- Clinical Yield: The retrospective NEJM AI study reported 18 diagnoses from 376 previously unresolved cases, a 4.8% additional yield after expert review.
- Workflow: Researchers gave the model de-identified phenotype terms, clinical notes, metadata, and filtered variant tables, then required evidence-linked explanations.
- Guardrails: Findings counted only after specialist review, ACMG/AMP-style variant assessment, CLIA confirmation, and clinical return of results.
- Operational Lesson: The useful pattern is periodic expert-led reanalysis, not autonomous diagnosis or direct patient decision-making.
Builder takeaway: Teams building clinical AI should copy the review loop: versioned prompts, source-linked hypotheses, auditable handoff, and explicit clinical confirmation gates.
Official source ->Ray Serve LLM on GKE Gets 5x Throughput
Google Cloud and Anyscale published Ray Serve LLM optimizations for GKE with up to 5x higher throughput and 8x lower latency.
- Performance: The updated stack claims up to 5x higher throughput and 8x lower latency compared with earlier Ray Serve configurations.
- Architecture: HAProxy routing, direct token streaming, and a v2 Ray executor backend move Ray away from hot-path token streaming bottlenecks.
- Hardware: Benchmarks used Google Cloud A4 VMs with NVIDIA HGX B200 systems and Gemma 4 E2B to isolate orchestration overhead.
- Version: Google points developers to Ray 2.56 and later for the improved Ray Serve LLM path on GKE.
Builder takeaway: Inference platform owners should retest Ray Serve before replacing it with raw vLLM, because the updated data path narrows the old flexibility-versus-speed tradeoff.
Official source ->Cloud Network Insights Adds Cross-Cloud Observability
Google Cloud made Cloud Network Insights generally available for hybrid and multi-cloud network path visibility with Broadcom AppNeta.
- Scope: The service watches paths across Google Cloud, on-premises environments, AWS, Azure, ISPs, and internet-facing applications.
- Telemetry: Active synthetic probing measures round-trip time, packet loss, jitter, and hop-by-hop path behavior even when user traffic is quiet.
- Automation: Policies can create and remove monitored paths from tags, while auto-baselining routes alerts to email, Slack, or PagerDuty.
- AI Assist: Gemini Cloud Assist can query Network Insights telemetry alongside broader infrastructure metrics for root-cause analysis.
Builder takeaway: SRE teams running multi-cloud apps can turn network debugging from reactive packet archaeology into monitored SLO evidence with path-level baselines.
Official source ->Google Cloud Details Europe's Digital Sovereignty Path
Google Cloud published its Europe digital sovereignty position, emphasizing compliance, open collaboration, and customer control.
- Positioning: Google frames sovereignty around customer choice, compliance controls, and collaboration rather than isolated national-only stacks.
- Timing: The post landed as European buyers pressure cloud vendors for clearer operational, legal, and data-residency assurances.
- Architecture: The practical work is policy enforcement, encryption, access transparency, workload placement, and partner-operated controls.
- Buyer Impact: Regulated teams should map sovereignty claims to audit artifacts, support boundaries, incident access, and exit planning.
Builder takeaway: Do not treat sovereignty as a marketing label; turn each vendor commitment into testable control evidence before production procurement.
Official source ->OpenAI GPT-5.4 AI Chemist Improves Chan-Lam Coupling
OpenAI and Molecule.one reported a near-autonomous chemistry workflow where GPT-5.4 improved a difficult medicinal chemistry reaction.
- Experiment: GPT-5.4 worked with Molecule.one's Maria AI and lab platform on Chan-Lam coupling for primary sulfonamides.
- Scale: Maria Lab ran 10,080 reactions for the OAI-M1-03 program across two cycles of high-throughput experimentation.
- Result: Optimized conditions improved measured yields for 88% of boronic acids and 83% of sulfonamides tested.
- Human Loop: Scientists selected proposals, corrected experimental details, handled basic lab operations, and validated representative bench-scale reactions.
Builder takeaway: The significant part is not a fully autonomous lab; it is a repeatable loop where models generate hypotheses and automation tests them quickly.
Official source ->Mandiant Tracks ShinyHunters PeopleSoft Zero-Day
Mandiant and Google Threat Intelligence linked UNC6240 ShinyHunters activity to exploitation of Oracle PeopleSoft CVE-2026-35273.
- Severity: Oracle PeopleSoft CVE-2026-35273 is a critical Environment Management component RCE with CVSS 9.8.
- Campaign: Observed exploitation ran from May 27 to June 9, before Oracle's June 10 advisory, making the activity zero-day exploitation.
- Targets: Google notified more than 100 potentially exposed organizations, with 68% of them in higher education.
- Defense: Mandiant recommends disabling or removing PSEMHUB, blocking exposed endpoints, and checking WebLogic and filesystem indicators.
Builder takeaway: PeopleSoft owners should treat externally reachable Environment Management Hub paths as an emergency exposure, not a routine patch queue item.
Official source ->NVIDIA France AI Factories Push European Compute
NVIDIA described a French AI infrastructure buildout spanning Mistral compute, Blackwell instances, Vera Rubin production, and open models.
- Mistral: NVIDIA says Mistral's Bruyeres-le-Chatel deployment is operational with 18,000 GB200 systems and a 200 MW European compute roadmap.
- Campus: Campus AI includes a planned 1.4 GW facility backed by Bpifrance, MGX, Mistral, and NVIDIA participation.
- Access: Scaleway now offers NVIDIA Blackwell B300-SXM instances for European developers and enterprises.
- Supply Chain: Bull and Foxconn plan European production and validation for NVIDIA Vera Rubin NVL72 systems.
Builder takeaway: European AI sovereignty is moving from policy language to power, chips, factory design, model ecosystems, and regional cloud capacity.
Official source ->This Week in Tech
Developer Resources
Key Takeaways
- 1Teams building clinical AI should copy the review loop: versioned prompts, source-linked hypotheses, auditable handoff, and explicit clinical confirmation gates.
- 2Inference platform owners should retest Ray Serve before replacing it with raw vLLM, because the updated data path narrows the old flexibility-versus-speed tradeoff.
- 3SRE teams running multi-cloud apps can turn network debugging from reactive packet archaeology into monitored SLO evidence with path-level baselines.
- 4Do not treat sovereignty as a marketing label; turn each vendor commitment into testable control evidence before production procurement.
- 5The significant part is not a fully autonomous lab; it is a repeatable loop where models generate hypotheses and automation tests them quickly.