News archive
October 17, 2026
Elorian’s Glimpse treats images as a workspace for reasoning
Elorian’s Glimpse offers API access at $0.60 input/$3 output per million tokens and claims strong chart and video results; the real question is repeatable accuracy on your own media.
0.60 USD per million input tokens3.00 USD per million output tokensfew hundred GPUs used for training, company reported
NVIDIA opens Nemotron 3.5 Super VL for multimodal agents
Nemotron 3.5 Super VL is open now with 120B total and 12B active parameters; NVIDIA claims up to 5× throughput, so a local or hosted trial can test cost against quality.
120B total parameters12B active parameters per request5× reported throughput gain
GPT-Live-1.1 adds 13 voices and new accents to ChatGPT
GPT-Live-1.1 adds 13 voices across French, German, Hindi, Japanese, Korean, Mexican Spanish, and English; voice quality improved, but OpenAI has not published language-specific accuracy rates.
13 new voices6 new language options
Claude finds the timestamp source behind NIST ECDSA seeds
Claude traced five ECDSA curve seeds to a timestamped lrand48 sequence in under 30 minutes; Anthropic says this explains their origin, not a break in the cryptography.
5 NIST curves explained<30 min reported search time$36,864 bounty donated
Ramp opens Glass, its internal AI workspace, to early users
Ramp’s Glass turns company knowledge into a shared AI workspace; the early-access story reports 323 contributors and 6,500 edits after a $1.5M single-week model bill.
323 employees reported as contributors6,500 reported knowledge edits$1.5M reported one-week model spend
OpenAI and Meta warn judges that deepfakes may outrun detection
OpenAI and Meta told a US judicial panel that deepfake detection may not scale to every case; evidence handling will need source records, expert review, and more than one signal.
White House AI task force starts a 120-day charter
The White House AI task force has a 120-day window, three named risk areas, and planned breach-reporting protocols; enforcement and legal authority remain unsettled.
120 days charter window3 named risk areas
OpenCode 2.0 joins terminal, desktop, and web work around one service
OpenCode 2.0 moves clients onto one local service and adds background subagents and async shell work, so builders can test parallel coding without changing models.
80+ internal plugins described3 client surfaces joined
Krea 3 makes image editing the main event
Krea 3 is an image editing model with native 2K output, a public beta, and a vendor comparison showing edits in under eight seconds at about $0.06 each.
<8 sec vendor edit example$0.06 vendor launch price per edit2736×1536 native output example
TwelveLabs launches Jockey for searching huge video libraries
Jockey ranked first on Product Hunt on October 16 and lets users query thousands of hours of video, returning cited moments and downloadable clip compilations.
372 Product Hunt points when checked#1 Product Hunt daily rank
Harness puts coding agents and machines in one open command center
Harness ranked second on Product Hunt with 291 points and offers a free open-source command center for agents across machines, with separate panes and branches.
291 Product Hunt points when checked49 non-coding harnesses listedfree Mac and Linux access
Konnect turns KiCad into an MCP-controlled PCB workspace
Konnect exposes 217 KiCad tools through one Rust binary, giving an AI client a way to inspect and change PCB work; its review and manufacturing checks still need human validation.
217 KiCad tools exposed10 KiCad version targeted
Mercor maps a large gap between AI benchmarks and real work
Mercor found only 208 of 923 occupations with enough benchmark prompts and estimates 90% of US working hours lack good coverage, so benchmark scores miss much of buyer work.
208/923 occupations with enough prompts90% working hours with little coverage6/22 job families with no such occupations
Goodfire argues neural networks may be more interpretable than assumed
Goodfire reports 99–100% single-feature behavior in a retrained vision model and a small unexplained share in Qwen3-32B; the site presents this as a hypothesis, not settled science.
99–100% reported single-feature behavior90% Qwen3-32B neurons with one main job, reported2–4% unexplained share, reported
A hardware reverse-engineering harness makes AI useful to driver work
A GNOME contributor used an AI agent plus USB trace capture to build a scanner driver in hours; the result worked only after human checks on real hardware.
300–7,200 reported scanner dpi range34.6 sec reported faster scan time54.4 sec reported earlier scan time
Navige launches a cost manager for running AI agents
Navige ranked second on Product Hunt on October 15 and offers a free Watchtower mode that previews model routing, caching, budgets, and stop controls before activation.
251 Product Hunt points when checked$9 paid-plan monthly minimum20% maker share of measured savings
