Frontiera Tech @FrontieraTechIT

Jev is not another chatbot. Diogo Almeida’s TypeSafe model takes state + typed questions and returns choices, scores or probabilities in one parallel pass. 70–500 ms, $0.042 per million input tokens, output free. Developers are already using it as the fast filter before an LLM writes anything. #jev #AI 👉 techcrunch.com/2026/09/18/a-n…

This is the future we shall bring into being

⭐ paperless-ngx/paperless-ngx github.com/paperless-ngx/… 📈 docling-project/docling github.com/docling-projec… 💎 woodpecker-ci/woodpecker github.com/woodpecker-ci/…

📡 GITHUB RADAR | 2026/09/20 ⭐ paperless-ngx/paperless-ngx Self-hosted document manager that scans, OCRs and indexes paper + PDFs into a searchable archive. Helps anyone ditch Evernote or Google Drive for private paperless storage. 📈 docling-project/docling Parses PDFs, Office files, images and more into structured Markdown or JSON ready for LLMs. Helps teams feed local models clean data without closed conversion services. 💎 woodpecker-ci/woodpecker Lightweight self-hosted CI/CD engine with plugins, low RAM use and simple YAML pipelines. Helps small teams run builds without the weight of Jenkins or full GitLab. Which one would you open first tonight? #GitHubRadar #OpenSource #SelfHosted Link in first reply 👇

Sources: 👇 1️⃣ @rasbt → 2️⃣ @AIpulse_is → x.com/AIpulse_is/sta… 3️⃣ @The420in → x.com/The420in/statu… 4️⃣ @ErSonusaini1 → x.com/ErSonusaini1/s… 5️⃣ Techmeme / SCMP coverage of Alibaba RADAR open-source drop (20 Sep)

Sebastian Raschka @rasbt

21 hours ago

It’s easy to dismiss Jev it as “just a classifier”. But people (me included) who have been training encoder-style models for classification for many years know they were usually special-purpose and limited in some way. The breakthrough of Jev is that it generalizes well. And

116 164 2K 188K 870

📡 Top AI News - 2026/09/20 🔥 1️⃣ TypeSafe’s Jev decision model keeps spreading among developers. Sebastian Raschka points out the real advance is generalization: it works across tasks where older encoder classifiers stayed narrow and brittle. 2️⃣ Sam Altman will brief the UN Security Council in person next week on AI and international security. The session comes weeks after frontier-lab leaders publicly backed a coordinated slowdown. 3️⃣ Trump said he will name an AI czar and stand up an “AI Force” modeled on Space Force. He called safety concerns a hoax and vowed not to slow U.S. development relative to China. 4️⃣ Anthropic is weighing a new model release before its planned IPO to blunt GPT-6 Astra’s enterprise gains. Reports also put its annualized revenue pace above $100 B. 5️⃣ Alibaba’s Damo Academy open-sourced RADAR, a medical vision model trained on 420 k+ CT exams that flags nearly 150 abdominal conditions. Hospitals can run it locally without per-scan API fees. #AI #MachineLearning #Agents

👉 energy.gov/science/articl…

Phase I gives early milestone awards up to $1.5 M per team. Phase II splits a $100 M pool for the 100-logical-qubit bar, plus two $50 M bonuses for 150 and 200. Private-sector companies only. Applications due 19 October. Webinar 25 September.

⚛️ DOE launched the Quantum Genesis Q Competition: up to $215 million for fault-tolerant, scientifically relevant quantum computers. Target: at least 100 logical qubits capable of hundreds of millions of fault-tolerant operations, plus real scientific demonstration programs. #quantum #genesis

@FrontieraTechIT CVE-2026-76460 being in KEV with active exploitation makes ISE an incident-response item, not a routine patch ticket. Prioritize the management interface and review admin/auth logs around the exposure window after applying Cisco’s fixed release.

🌦️ WeatherNext 3 di Google DeepMind è il loro modello meteo globale più avanzato. Aggiorna le previsioni ogni ora usando dati satellitari in tempo reale. Risoluzione fino a 5 km per temperatura e umidità, 10 km per vento e precipitazioni. Miglioramenti misurati su precipitazioni e variabili di superficie rispetto alle versioni precedenti e a modelli numerici tradizionali. #google #meteo #ai 👉 blog.google/innovation-and…

‘Stomach bugs’ send protective immune cells rushing from gut to brain by Gemma Conroy @Nature Learn more: bit.ly/4h8XDro #Neuroscience #MedTech #HealthTech #Tech #TechForGood

👉 deepmind.google/blog/alphageno…

They added an AlphaGenome Variant Impact (AVI) score that combines coding and non-coding effects into one number. Researchers can query it through a website, API, or as a skill inside Google’s tools. Early uses include prioritizing variants in unsolved rare-disease cases and spotting more non-coding associations in UK Biobank data.

🧬 Google DeepMind released AlphaGenome Atlas: predictions for every possible single-letter DNA change in the human genome. That’s 9 billion variants. The dataset is about 1 petabyte. #Google #AI

🚨 ATTENTION: Public coding leaderboards keep climbing. Real-SWE just asked a colder question: can a frontier agent finish the kind of ticket a salaried engineer actually gets inside a company it has never seen? Specific Labs licensed private production codebases — a consumer app with 200K+ users, a fintech stack that processed 100K+ bank statements, enterprise sales workflows — and turned real engineer work into 10 tasks. Models got eight independent runs each, scored pass@1, and were evaluated inside their native harnesses rather than in a lab vacuum. Fable 5.1 + Claude Code leads at 38.8%. GPT-6 Astra + Codex CLI is at 33.8%. Grok 4.6 + Grok Build lands at 32.5%. Gemini 3.8 Flash is 31.2%. GPT-5.6 Sol finishes last at 16.2%. Six of the ten tasks resolve below 15%. Nobody solved the analytics stream reducer across 64 attempts. The failure modes are painfully familiar to anyone who has onboarded a new hire: missed hidden requirements, integration errors, unverified assumptions about how tax, billing, identity, or metering already works. Median reference solutions touch 11 files. The prompt is underspecified on purpose, the way a real ticket is. The useful takeaway is not “agents are useless.” It is that solving a contaminated public puzzle is not the same job as navigating company-specific conventions with business consequences. Would you ship agent-written billing or tax changes without a human merge gate? 👉 withspecific.com/benchmarks/rea… thenewstack.io/real-swe-codin… #AI #CodingAgents #DevTools

Sources & further reading: • The Hacker News → thehackernews.com/2026/09/cisco-… • CISA / The Hacker News → thehackernews.com/2026/09/cisa-f… • The Hacker News → thehackernews.com/2026/09/critic… • BleepingComputer → bleepingcomputer.com/news/security/… • BleepingComputer → bleepingcomputer.com/news/security/… #InfoSec #ThreatIntel

🛡️ CYBER BULLETIN | 2026/09/20 🚨 1. Cisco ISE CVSS 10 auth bypass under active attack CVE-2026-76460 lets unauthenticated attackers bypass the management interface on Identity Services Engine and ISE-PIC. Cisco confirmed exploitation; CISA added it to KEV with a short remediation window. Patch to the fixed releases immediately. 2. CISA flags three Linux kernel flaws as actively exploited CVE-2025-39682, CVE-2026-53266 and CVE-2025-39964 went into the KEV catalog. Federal agencies must remediate by September 21. Red Hat confirmed public exploits exist for at least one of them. 3. Orkes Conductor pre-auth RCE exploited in the wild CVE-2026-58138 (CVSS 9.8) allows unauthenticated remote code execution via malicious workflow definitions. Fortinet blocked nearly 7,000 attempts in one week. Upgrade to 3.30.2 or later. 4. Gyazo breach exposes 23.6 million user records Attackers exploited an image-upload server flaw on September 11 and stole account data plus 490 million image metadata records. Password hashes, emails and some tokens are among the exposed items. 5. ShinyHunters defaces Clop ransomware leak site The group claimed full access after an unauthenticated Grav CMS upload, replaced the Tor site with its own branding and said it plans to extort the rival operation. #Cybersecurity #CISA #NIST

🛡️ CYBER BULLETIN | 2026/09/20 🚨 1. Cisco ISE CVSS 10 auth bypass under active attack CVE-2026-76460 lets unauthenticated attackers bypass the management interface on Identity Services Engine and ISE-PIC. Cisco confirmed exploitation; CISA added it to KEV with a short remediation window. Patch to the fixed releases immediately. 2. CISA flags three Linux kernel flaws as actively exploited CVE-2025-39682, CVE-2026-53266 and CVE-2025-39964 went into the KEV catalog. Federal agencies must remediate by September 21. Red Hat confirmed public exploits exist for at least one of them. 3. Orkes Conductor pre-auth RCE exploited in the wild CVE-2026-58138 (CVSS 9.8) allows unauthenticated remote code execution via malicious workflow definitions. Fortinet blocked nearly 7,000 attempts in one week. Upgrade to 3.30.2 or later. 4. Gyazo breach exposes 23.6 million user records Attackers exploited an image-upload server flaw on September 11 and stole account data plus 490 million image metadata records. Password hashes, emails and some tokens are among the exposed items. 5. ShinyHunters defaces Clop ransomware leak site The group claimed full access after an unauthenticated Grav CMS upload, replaced the Tor site with its own branding and said it plans to extort the rival operation. #Cybersecurity #CISA #NIST

Check out DiffusionGemma as Jev. It’s even faster on several tasks. Credit @mmastrac

Google Gemma @googlegemma

3 days ago

"DiffusionGemma as Jev" showcases the power of non-autoregressive architectures. While Jev demonstrates the value of rapid decision models, running DiffusionGemma in this paradigm leverages canvas diffusion to evaluate structured choices in a single parallel pass: ⚡ ️Massive

55 501 4K 728K 3K