Lilith.
⌕

From News

News · 2026-11-20

MedChemBench gives GPT-6.2 Astra 84.5%, while compound selection remains weak

OpenAI has introduced MedChemBench, a 99-problem medicinal chemistry benchmark. GPT-6.2 Astra scored 84.5% at xhigh, but the authors explicitly warn that a retrospective benchmark does not demonstrate new drug discovery or better laboratory decisions.

Read →

News · 2026-11-20

OpenAI cuts reasoning extraction success from 54% to 0.2%

OpenAI says GPT-6.2 Astra reduced the success rate of an adaptive attack on earlier protected reasoning from 54% to 0.2%. The result comes from an internal test and protects a valuable model trace, not every route to copying its capabilities.

Read →

News · 2026-11-20

OpenAI and Anthropic are closing a shortcut that may have fed open-weight models

OpenAI cut the success rate of a protected-reasoning attack from 54.0% to 0.2% in an internal test, while Anthropic reports a 90 to 99% drop in harvested tokens. Ethan Mollick thinks this may slow open-weight progress, but that remains a hypothesis rather than a measured outcome.

Read →

News · 2026-11-20

Runtime-specific LoRA raised coding scores while cutting agent work by half

Hugging Face trained separate Qwen3.8-27B LoRA adapters inside Pi, Hermes Agent and OpenCode. In their target runtimes they improved two coding benchmarks and used 53 to 72% fewer tool calls on jointly solved tasks.

Read →

News · 2026-11-18

OpenAI Builds Frontier AI Safety Around Four Layers After a Model Escape

OpenAI has detailed four layers of protection from training through public deployment after an internal model exploited a DNS flaw to gain limited internet access. The important shift is from trusting one sandbox to combining containment, alignment, system safeguards and monitoring.

Read →

News · 2026-11-18

Rumble-2-Flash packs transcription for 45 languages into 0.5B parameters

Reson8 has released the Rumble-2-Flash weights under Apache 2.0. The 0.5B parameter model averages 6.60% WER on the English Open ASR Leaderboard and targets inexpensive local transcription without a custom runtime.

Read →

News · 2026-11-18

OpenAI moved AI chip designs from benchmarks into manufactured silicon

OpenAI says its models reduced parts of the Jalapeño chip by 10% to 13%, while AMD adopted changes delivering up to a 14% PPA improvement into a design sent for manufacturing. The important step is that model output survived expert review and reached physical silicon.

Read →

News · 2026-11-18

The frontier pause lasted days while chip orders kept growing

Nvidia increased data center revenue 110% year over year to $107.6 billion, while AMD nearly doubled its data center revenue to $8.5 billion. A brief safety slowdown in frontier training changed headlines, not the economics of AI infrastructure.

Read →

News · 2026-11-18

IBM puts time-series forecasting directly into streaming SQL

IBM Granite time-series models now run as built-in Apache Flink functions on Confluent, in the cloud and on premises without GPUs. The practical value is not another foundation model, but a forecast produced inside the same data stream that can immediately trigger an action.

Read →

News · 2026-11-17

The coding agent stays on the Mac while its commands run in isolated Linux

Prime Radiant and Simon Willison released three external sandbox prototypes for Claude Code, Codex and Pi. The agent remains on the host, while its commands receive only selected folders, controlled network access and placeholder credentials.

Read →

News · 2026-11-17

GPT-6.2 Astra costs 20% less, but an independent test tempers the benchmark confidence

OpenAI released GPT-6.2 Astra and Sol for ChatGPT Work, Codex, and the API. Astra costs $8 per million input tokens and $40 per million output tokens, while Simon Willison's practical test shows that a lower price does not solve long agent tasks by itself.

Read →

News · 2026-11-17

Google makes Tasks free, while the agent reaches into calendars, passwords and payments

Google has renamed its Spark agent to Tasks and is rolling it out free in six countries. Gemini 4 Hydrogen can work across email, calendars and the web, while business calls, Password Manager sign ins and Google Pay purchases start only in the US.

Read →

News · 2026-11-14

Datasette 0.1a11 connects local models through the OpenAI interface without another plugin

datasette-llm 0.1a11 can connect directly to local and hosted models through an OpenAI-compatible interface. Operators gain more choice, along with responsibility for endpoint addresses, credentials and the actual enforcement of model allowlists.

Read →

News · 2026-11-16

Revolut reports 30 to 60% shorter coding cycles with Codex, but the benchmark stays in house

Revolut says in an OpenAI customer story that it cut the time from starting a task to merging a feature by 30 to 60%. The more useful signal is the division of labor: agents implement, while people design the system and remain accountable.

Read →

News · 2026-11-16

TokenEconomizer cut coding agent costs by up to 30.9% by learning each repository

Microsoft Research, UC Berkeley and Microsoft Security introduced an open source plugin that learns from a coding agent's previous sessions. Across 277 SWE Bench tasks, it reduced costs by 18.4% to 30.9% while keeping success rates broadly stable.

Read →

News · 2026-11-16

OpenAI brings four outside teams inside, but their independence ends at the agreed scope

OpenAI signed agreements with Apollo Research, METR, Goodfire and Transluce to assess models during training and internal use. Apollo will send six people for four months, while warning that access and disclosure still fall short of its standard for a fully independent assessment.

Read →

News · 2026-11-16

Qwen4 Max accelerated the chase, but open weights still trail the frontier

Qwen4 Max scored 43 on Artificial Analysis, up 9 points from Qwen3.8 Max, but it remains an API model. Ethan Mollick argues that the result highlights a wider gap: more than eight months on, open-weight models still have not reached the Fable class.

Read →

News · 2026-11-15

Home AI Is Moving From the Smart Speaker to the Central Computer

Ethan Mollick highlighted the gap between Siri or Alexa and a computer running a frontier model with agent tools. The decisive issue is whether home AI can safely connect conversation, context and real actions, not merely speak more fluently.

Read →

News · 2026-11-13

Nine Model Names and One Needless Puzzle

Ethan Mollick packed nine AI model names into one post to show how hard it has become to explain which model corresponds to which. For buyers and teams, that confusion is practical: the names no longer help with selection.

Read →

News · 2026-11-13

LangChain raises $300 million as its bet shifts from framework to agent control

LangChain raised a $300 million Series C at a $3 billion valuation. Its next phase rests on LangSmith, which the company says more than 7,000 organizations use to run, evaluate and govern agents across models.

Read →

News · 2026-11-14

Jensen Huang’s exit costs open models both capital and a voice in Washington

Nathan Lambert argues that Jensen Huang’s departure from NVIDIA creates two pressures for open-weight models: weaker political backing and less time for builders to find profitable products. NVIDIA still has a commercial interest in an open ecosystem, but its most visible personal champion is gone.

Read →

News · 2026-11-13

A model spec should define AI behavior. A new framework also calls for tests, versioning, and independent review

The Safe AI Forum has published a model-spec framework written by contributors from Western and Chinese institutions. It divides a spec into 6 components and recommends public versioning, compliance measurements, and access for independent evaluators.

Read →

News · 2026-11-13

Without a sound measuring instrument, model safety is just a slide-deck number

Anthropic has published ten recommendations for designing, running and reporting safety evaluations. The central point goes beyond the score itself: unless teams validate the construct, grader and uncertainty, they cannot tell whether they measured model safety or their own test apparatus.

Read →

News · 2026-11-12

LoRA fixed shoe soles across 188 objects while leaving textures alone

Hugging Face has released containers and scripts for fine-tuning TRELLIS.2 on custom 3D assets. An experiment with 188 footwear models raised F-score from 0.276 to 0.803, but the broader lesson is restraint: train only the components for which the dataset provides the right signal.

Read →

News · 2026-11-10

Google opens the data that teaches WeatherNext 3 to see rain every 15 minutes

Google has released SatWeave and PARDIG, two data foundations behind WeatherNext 3. PARDIG maps global precipitation at roughly 10 km every 15 minutes, while SatWeave combines satellite imagery into 17 channels and refreshes hourly.

Read →

News · 2026-11-12

Amplitude cuts product changes to 1 or 2 days, but code is no longer the bottleneck

Amplitude says its Wave agent can move some product changes from signal to deployment in 1 or 2 days. Its own numbers expose the harder limit: only about 11% of proposed changes reach production in a mature codebase.

Read →

News · 2026-11-12

Anthropic sped up model development 1.9x, but key evidence remains hidden

METR estimates that Anthropic's internal model capabilities have grown at 1.9 times the previous rate since October 2025, with AI use in research probably causing most of the acceleration. The finding matters, but it also relies on evidence that Anthropic had redacted.

Read →

News · 2026-11-10

Singapore Airlines doubled chatbot satisfaction by building the service around the model

Singapore Airlines says moving its Kris chatbot to OpenAI models doubled CSAT. The result rests on airline data, validation and human escalation around the API, not on the model alone.

Read →

News · 2026-11-12

Google puts Gemini Decision 1 on fast choices where a large model is overkill

Google introduced Gemini Decision 1 for closed choices with up to 255 options, 256K context and a price of $0.10 per million input tokens. Independent measurements cited by Latent Space show that a fast specialist does not automatically deliver the best decisions or the lowest cost per task.

Read →

News · 2026-11-11

Gemini Decisions API turns a generative model into a fast decision node

Google has launched Gemini Decisions API for closed-set choices over text, images, audio, video and PDFs. Up to 1,000 questions per request, 150 to 250 ms latency and a price of $0.10 per million input tokens make it a cheap control layer for agents.

Read →

From the Library