Spectrum Web Co

Spectrum Web Co is our high-impact software development agency specialising in secure, scalable dApps and cloud-native infrastructure. We build with precision.

[Spectrum Web Co @spectrumwebco]

OpenAI rebuilt the Git backend behind its monorepo because Codex agents were swamping it. Pando handles each branch independently and reuses packs across clients, cutting median push latency nearly 5x. Built in Rust in 10 weeks.

https://t.co/Bvd9uqA0CM

#AI #DevTools

[Spectrum Web Co @spectrumwebco]

Trajectory raised a $45M Series A at a $300M valuation, led by Sequoia with Nvidia in. It lets lawyers, analysts and support teams keep post-training open models on real production feedback, each update gated by their own evals.

https://t.co/T0GdoXaUAZ

#AI #OpenSource

[Spectrum Web Co @spectrumwebco]

The FBI says North Korea's TraderTraitor was behind the record $6.3B Binance hack. ~$3B was frozen on-chain, withdrawals reopened with $3.64B net outflow on day one, and Binance published a zero-knowledge proof of reserves for 10 tokens.

https://t.co/h2llc2gtRM

#Web3 #Crypto

[Spectrum Web Co @spectrumwebco]

Marvell is buying Lumentum in an all-stock deal at a $115B implied value, betting AI data centers are moving from copper to optics. Lasers, transceivers and the chips that drive them under one roof, with closing expected in 1H 2028.

https://t.co/m2gH6oPYf5

#AI #Semiconductors

[Spectrum Web Co @spectrumwebco]

AMD just had its biggest quarter ever: $13.4B revenue, up 45%, with data center up 97% to $8.5B. Anthropic, Meta, Microsoft, OpenAI and Oracle are running on its first Helios racks, and it's guiding to ~$17B next quarter.

https://t.co/Rf1ONJ0BW2

#AI #Semiconductors

[Spectrum Web Co @spectrumwebco]

Hugging Face's Carbon-X is an open 14B model you can chat with about DNA. It nearly matches GPT-6.1 Sol on DNA benchmarks at ~1/1,800 the cost per problem, and runs on one 80 GB GPU so sensitive sequences stay local.

https://t.co/0ELT7iw7xl

#AI #OpenSource

[Spectrum Web Co @spectrumwebco]

Anthropic open-sourced FlashProve, a zero-knowledge proof that a reply came from the model you were promised, without revealing its weights. On a 400B model, one H100 proves ~110 prefill tokens/sec; 19 MB proofs check on a CPU in ~17s.

https://t.co/dtHAGpdNuC

#AI #OpenSource

[Spectrum Web Co @spectrumwebco]

Google DeepMind gave its Antigravity agent RDKit-based cheminformatics skills. In a lung cancer demo it mined thousands of compounds and proposed new drug designs in about an hour.

https://t.co/3LaRnPZvNo

#AI #DrugDiscovery

[Spectrum Web Co @spectrumwebco]

Fireworks used its first Forge conference to launch Ultra-fast inference at up to 1,500 tokens/sec, a half-price Background Mode for agents, and Sovereign, which runs its engine on your own GPUs.

https://t.co/DNtWwj27GS

#AI #OpenSource

[Spectrum Web Co @spectrumwebco]

OpenAI says GPT-6 Astra designed a working Codex gadget, enclosure, circuit board and firmware, in under a week. It still missed factory tolerances. OpenAI is now teaming with Mazda and Chip Ganassi Racing on engineering.

https://t.co/8FCQzR5MoG

#AI #OpenAI

[Spectrum Web Co @spectrumwebco]

Apple just opened Siri AI to every app in the iOS 27.3 betas. App Intents become actions Siri can call, and new app skills, written as SKILL.md files, chain multi-step requests. Customers get it in a few months.

https://t.co/fznCJqb3cQ

#AI #Apple

[Spectrum Web Co @spectrumwebco]

OpenAI's GPT-Rosalind-6 is GPT-6 Astra's weights with fewer refusals on dual-use biology, offered only to vetted labs, companies and governments. OpenAI says it beats Rosalind-5.5 on med chem (49.6% vs 28.9%) and genomics (37.7% vs 16.7%).

https://t.co/75tCuqwUzH

#AI #Biotech

[Spectrum Web Co @spectrumwebco]

NC AI built VAETKI Vision 3B, a tiny vision-language model for robots and factory gear. It says it beats 4B and 8B Qwen and Ministral models on spatial reasoning, and a welding fine-tune nearly doubles weld classification scores.

https://t.co/pZJuf7sMv1

#Robotics #AI

[Spectrum Web Co @spectrumwebco]

Flower Labs just shipped Endeavor 1.1: image input, inference on EU-only GPUs, and raw thinking traces you can read and train your own models on. Benchmarks are self-reported, so test before trusting.

https://t.co/SBPk1wwV8w

#AI #SovereignAI

[Spectrum Web Co @spectrumwebco]

AGIBOT unveiled GO-3 Preview, a robot foundation model it says ranks first on RoboDojo, RoboTwin and RoboColiseum. Plus a 100-robot practice fleet and an agent layer so deployed robots keep learning. Open source promised by end of 2026.

https://t.co/qA2dp9deUs

#Robotics #AI

[Spectrum Web Co @spectrumwebco]

Anthropic's amended S-1 adds Q3: $16.96B revenue, up 1,127% YoY, though QoQ growth slowed to 47%. Net loss $23.3B, mostly fair-value charges. And ~$895B in committed spend, nearly all compute. Listing on Nasdaq as ANT.

https://t.co/X9cHvRx1s8

#AI #Anthropic

[Spectrum Web Co @spectrumwebco]

SemiAnalysis launched GDPx: 206 hours-long agent tasks across 11 domains like chip design, quant trading and cyber. The best model, Claude Opus 5.5, scores just 12.0%. Gemini 4 Carbon gets 10.9% at about a third of the cost per task.

https://t.co/SmmzQmZf6E

#AI #Agents

[Spectrum Web Co @spectrumwebco]

OpenAI's new LifeVal tests whether AI agents can run everyday errands: picking a kindergarten, meal plans, paperwork. Reported average reward went from 49.3% (GPT-5.2) to 80.6% (GPT-6 Astra), but the tasks and grading aren't public.

https://t.co/hG6s8qdPEW

#AI #Agents

[Spectrum Web Co @spectrumwebco]

Anthropic put message threads into public beta on the Claude API: agents send a conversation once, then only new turns. Anthropic says median time to first token fell ~35% on 500k–1M token requests. Same prices and rate limits.

https://t.co/8qZWn6rmM7

#AI #Claude

[Spectrum Web Co @spectrumwebco]

ServiceNow open-sourced Apriel Voice 27B Preview (MIT): one Gemma 4-based model that hears speech, reasons and makes tool calls, with TTS left swappable. 65.5% τ-Voice task success at 1.9s latency.

https://t.co/psMkKSqrOX

#AI #OpenSource