Skip to main content
Back

Anthropic's Claude Science runs 60+ databases: researchers report 10x

Anthropic launches Claude Science, NVIDIA splits 30B model for 2.42x speed, GitHub releases spec-kit, 35B model beats trillion-param giants

July 3, 2026 6 min read

Hey,

Anthropic had a big week. Claude Fable 5 is back after a 19-day security timeout, and Claude Science just launched in beta. Two very different stories, same underlying theme.

The real signal: Anthropic is quietly moving from model company to vertical platform company. Science for researchers, Code for engineers, tighter safety rails for governments. That's a product suite, not a chatbot.

Meanwhile, a 35B model just beat trillion-parameter giants on agent tasks. Bigger is no longer the only game.
In Partnership with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 6 min 47 sec

Top News

▸ Anthropic brings Claude Fable 5 back globally with tighter cybersecurity blocks

Lambda

▸ How to push Model FLOPS Utilization past 60%

Top News

▸ Anthropic launches Claude Science app with 60+ research databases and live code

Wispr Flow

▸ Voice-First Prompting for Devs

Top Model

▸ NVIDIA splits a 30B model in two to generate text 2.42x faster with 98.7% quality

Signals

1. Nous Research ships Hermes Agent v0.18.0 with new judgment capabilities
2. Presented by Viktor: Take the 4th off. Viktor, the AI employee, keeps the work moving across 3,000+ tools. Start now with $100 in credits
3. LangBot ships open-source platform to deploy AI bots across Slack, Discord, and WeChat
4. New 35B model beats trillion-parameter giants on long-horizon agent tasks
5. LiquidAI refreshes its tiny 450M vision-language model with stronger real-world performance
6. GitHub releases spec-kit, an open-source tool that structures AI coding before any code is written
Top News
Anthropic brings Claude Fable 5 back globally with tighter cybersecurity blocks
43,107 Likes
Anthropic brings Claude Fable 5 back globally with

Claude Fable 5 is back online after a 19-day suspension. Here is the quick version of what happened and what it means for you.

Amazon researchers found a way to bypass Fable 5's safety filters, prompting the model to identify software vulnerabilities and, in one case, produce code showing how to exploit one. The US government treated that as a national security issue and pulled the model globally.

To get it back, Anthropic made three changes:

  • A new safety classifier trained specifically on the reported bypass technique, blocking it in over 99% of cases.
  • Some routine coding and debugging requests will temporarily fall back to Opus 4.8 while Anthropic tunes the filters to cut false positives.
  • A new cross-industry jailbreak severity framework is being drafted with Amazon, Microsoft, Google, and other Glasswing partners.

If a classifier fires on your request, the response comes from Opus 4.8. Your code path stays the same. Access is live across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork.

Read more
forward →
Presented by Lambda
How to push Model FLOPS Utilization past 60%

Most large-scale training runs operate at 35–45% Model FLOPS Utilization (MFU). You're paying for twice the compute you're using.

Lambda's engineers benchmarked Llama 3.1 models from 8B to 405B on NVIDIA Blackwell GPUs and traced efficiency loss to its root causes:

  • Memory overhead
  • Parallelism strategy
  • Serialized communication

The result: a reproducible framework that pushed MFU past 60% (25%+ improvement vs. industry benchmark) with no changes to model architecture.

Get the guide
partner with us →
Top News
Anthropic launches Claude Science app with 60+ research databases and live code
37,375 Likes
Anthropic launches Claude Science app with 60+ res

Anthropic just shipped Claude Science, a dedicated research app now in beta. Think of it like a smarter Jupyter Notebook that actually runs your analysis, not just talks about it.

The big idea: general AI assistants can discuss biology, but they can't run a pipeline, navigate scientific databases, orchestrate cluster jobs, or keep track of what happened in a previous session. Claude Science fixes that.

Here's what you can actually do with it:

  • Connect natively to 60+ scientific databases, including UniProt, PDB, and ChEMBL
  • Every artifact includes the exact code that generated it, the environment it ran in, and the conversation that led there
  • A reviewer agent inspects outputs, flagging incorrect citations and figures that don't match their underlying code
  • Scale analyses from a single GPU to hundreds as needed

One UCSF team reported certain analyses now take roughly one tenth of the previous time. Available on macOS and Linux for Pro, Max, Team, and Enterprise plans.

Read more
forward →
Presented by Wispr Flow
Voice-First Prompting for Devs

Wispr Flow is the voice input layer for your AI workflow. Speak detailed prompts into Cursor, Claude, and ChatGPT in 20 seconds instead of typing them in 2 minutes.

Clean, technical text in any app on Mac, Windows, iOS, and Android. 89% sent with zero edits.

Used by teams at OpenAI and Vercel.

Try wispr flow
partner with us →
Top Model
NVIDIA splits a 30B model in two to generate text 2.42x faster with 98.7% quality
4,395 Likes
NVIDIA splits a 30B model in two to generate text

Normal AI models write text one word at a time, like typing letter by letter. NVIDIA just flipped that idea with Nemotron-Labs-TwoTower, a model that writes whole chunks of text in parallel.

Here is the trick: they took one 30B model and split it into two copies. One half reads your prompt and holds all the context. The other half takes a block of blank "masked" slots and fills them all at once, refining its guesses over a few passes until it is confident enough to commit. No training from scratch needed.

The results are pretty solid:

  • 2.42x faster generation vs. the original model
  • 98.7% quality retained across benchmarks like MMLU, GSM8K, and HumanEval
  • Runs on 2x H100 or A100 GPUs, about 59GB per GPU

You can load it today via Hugging Face Transformers with trust_remote_code=True and call generate_mask_diffusion() to use the fast parallel mode. This opens up faster inference for any text generation task without sacrificing much quality.

Read more
forward →
Signals
1
Nous Research ships Hermes Agent v0.18.0 with new judgment capabilities
2,584 Likes
2
Take the 4th off. Viktor, the AI employee, keeps the work moving across 3,000+ tools. Start now with $100 in credits
Presented by Viktor
3
LangBot ships open-source platform to deploy AI bots across Slack, Discord, and WeChat
16,603 Stars
4
New 35B model beats trillion-parameter giants on long-horizon agent tasks
619 Likes
5
LiquidAI refreshes its tiny 450M vision-language model with stronger real-world performance
771,000 Downloads
6
GitHub releases spec-kit, an open-source tool that structures AI coding before any code is written
1,718 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 250,000+ AI developers?
Work with us