Skip to main content
Back

OpenAI Astra solves 10 unsolved math proofs for $2,000 total

OpenAI model solves 10 math problems for $2k, Unsloth ships Qwen3 27B on 17GB, MiniMax releases H3, Google links AI consciousness to moral values

August 4, 2026 6 min read

Hey,

AI just cracked 10 open math problems for $2,000. Not a benchmark. Not a leaderboard. Actual unsolved theorems, some untouched for 30 years.

Meanwhile, Google quietly dropped a paper showing that when you train a model to deny having feelings, it loses moral reasoning too. Safety training that suppresses consciousness also suppresses empathy. You can't lobotomize one without the other.

The pattern? We're optimizing models in ways we don't fully understand yet.
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 34 sec

Top News

▸ OpenAI's unreleased model cracks 10 open math problems for $2,000 in compute

Unblocked

▸ [Webinar] Can you prove AI is working?

Top News

▸ OpenAI rebuilds ChatGPT Voice so reasoning runs without interrupting speech

Granola

▸ Always be Prepared, Even for Back-to-Back Meetings

Top Repo

▸ Unsloth ships day-zero support for Qwen3 27B running on 17GB RAM

Signals

1. Open-source tool lets Claude Code edit your videos automatically, no timeline needed
2. ngrok: Point Cursor, Zed, or any OpenAI-compatible coding agent at models running on any machine via ngrok
3. Open-source self-rewriting AI agent keeps memory and identity across restarts
4. Google study finds suppressing AI consciousness claims also kills its moral values
5. New paper proposes replacing prompt engineering with agent graph engineering
6. MiniMax releases its H3 model to the public
Top News
OpenAI's unreleased model cracks 10 open math problems for $2,000 in compute
10,791 Likes
OpenAI's unreleased model cracks 10 open math prob

OpenAI's next model, Astra, just solved 10 open math problems that nobody had cracked in over a decade. That's the debut.

Astra did not just guess answers. Humans used the model to write up the results, then Astra formalized each proof in Lean, a language for machine-verified theorems. Think of Lean as a strict computer checker that confirms the math is airtight.

The solved problems include:

  • The first-ever explicit construction of a non-sofic group, resolving a central question in group theory open since 1999.
  • A disproof of Connes's rigidity conjecture, new bounds for high-dimensional sphere packing, and results resolving several problems posed by Paul Erdos.

The cost? About $2,000 total at Sol's API rates. That is $200 per solved problem.

OpenAI published a 249-page manuscript collection, model-written reasoning walkthroughs, and Lean 4 certificates for all 10 results so you can read exactly how it got there.

Read more
forward →
Presented by Unblocked
[Webinar] Can you prove AI is working?

AI is in your engineering workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem. 

Join live on Aug 19 (FREE) to see:

  • The 4 metrics to measure the gap and what's capping your metrics
  • The 8 stages of context maturity and a free tool to pinpoint where your team is
  • Why more MCPs and bigger context windows aren’t enough, and what it takes to get real value from your agents.
Register now
partner with us →
Top News
OpenAI rebuilds ChatGPT Voice so reasoning runs without interrupting speech
7,051 Likes
OpenAI rebuilds ChatGPT Voice so reasoning runs wi

Old ChatGPT Voice had a clunky problem: it waited for you to stop talking, then thought, then replied. One side at a time. That awkward pause? That was the system guessing when your turn ended.

GPT-Live fixes this. It uses a full-duplex architecture, meaning it listens and speaks at the same time, just like a real conversation. It removes the old turn detector entirely, so the model can listen and speak simultaneously.

Here is what that unlocks:

  • Hard questions get quietly handed to GPT-5.5 in the background, while the conversation keeps flowing.
  • Live translation becomes possible because the model does not need a complete sentence before generating output.
  • You pick reasoning depth: Instant, Medium, or High, each routing to a different GPT-5.5 tier.

Session startup dropped from six network round trips to one, making it feel fast from the first word.

Read more
forward →
Presented by Granola
Always be Prepared, Even for Back-to-Back Meetings

Granola introduces Briefs - a meeting prep feature that automatically gathers context.

Briefs pull together who you're meeting, what you discussed last time, recent company news, and any relevant email threads.

Before the call, you simply glance at two or three lines of prep and join knowing the full context.

Try granola today
partner with us →
Top Repo
Unsloth ships day-zero support for Qwen3 27B running on 17GB RAM
6,857 Likes
Unsloth ships day-zero support for Qwen3 27B runni

Alibaba just dropped Qwen3.8-27B as open weights, and it fits on hardware you probably already own.

The big deal? 17GB of RAM or VRAM is all you need to run a 27-billion-parameter model locally. That's a single RTX 4090, a Mac with 24GB unified memory, or even RAM + VRAM combined. No datacenter required.

And you won't have to wait for tooling to catch up. Unsloth has day-zero support for both running and fine-tuning Qwen3.8. Fine-tuning means teaching the model new tricks on your own data, and Unsloth makes that 2x faster with 70% less memory than standard approaches.

Here's what this opens up for you:

  • Run a frontier-class model fully offline, no API costs
  • Fine-tune it on your own dataset right from a Colab notebook
  • Use it for coding, reasoning, and multimodal tasks including images and video
  • Switch between thinking mode (slow, deep reasoning) and fast chat mode in one model

Bottom line: a powerful local AI just got accessible to anyone with a decent GPU.

Read more
forward →
Signals
1
Open-source tool lets Claude Code edit your videos automatically, no timeline needed
4,059 Likes
2
Point Cursor, Zed, or any OpenAI-compatible coding agent at models running on any machine via ngrok
Presented by ngrok
3
Open-source self-rewriting AI agent keeps memory and identity across restarts
909 Stars
4
Google study finds suppressing AI consciousness claims also kills its moral values
1,938 Likes
5
New paper proposes replacing prompt engineering with agent graph engineering
1,022 Likes
6
MiniMax releases its H3 model to the public
3,546 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 300,000 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 300,000+ AI developers?
Work with us