Luke

[Lumina @LuminaXspace]

🚨Grok appears to be testing stock tickers built into search suggestions:

I have never seen this before in the search suggestions and I can't replicate it with any other combination or letter apart from "f" for some reason?

I presume it will be like smart cashtags on X possibly?

[Lumina @LuminaXspace]

🚨Claude Fable 5 has solved a research mathematics problem dating back to 1982

A lot of unsolved research maths problems are labour intensive, messy and require a lot of time which is why professional mathematicians hadn't produced a solution, but this is a perfect use of AI.

• This is the second open problem solved by Fable 5
• First elicited by David Roe in 1982
• This was also considered publishable research, this makes it a meaningful contribution not some bs.

We are quickly approaching the Singularity!

[Lumina @LuminaXspace]

🚨There have been two new OpenAI model names spotted

"Zinc" and "Magnesium" were apparently seen in DesignArena but they are not enabled yet.

• They are probably refreshed GPT 5.6 variants
• The "Zinc" one might be an update to Sol
• And "Magnesium" most likely an update to Terra

Tibo also leaked something about "performance and efficiency improvements" which would make sense.

This could be a nice upgrade to challenge Claude Opus 5 before we get GPT 6 in a few weeks if their Washington trip goes well.

https://t.co/hrJP8FxWPs

[Luke @luke_0302]

@lzccrr @LuminaXspace @pangram How is public info AI generated 💀

[Lumina @LuminaXspace]

🚨Fable 5.1 may be getting held back

Anthropic could already have its next flagship ready, but may be waiting for OpenAI to move first:

• GPT 6 and Fable 5.1 both drop in August
• Fable 5.1 should keep the same pricing as Fable 5
• I’m predicting they launch a few days to a week apart

There is a pattern of companies holding stronger models in reserve so they can answer a rival launch immediately and take away some of its impact.

Would you rather see Fable 5.1 now, or have both models drop back to back?

[Lumina @LuminaXspace]

🚨Kimi K3 open weights are live

Moonshot has released its enormous new frontier model:

• 2.8T-parameter MoE
• Native vision
• 1M-token context
• 16 of 896 experts active per token
• Around 2.5× better scaling efficiency than Kimi K2
• New attention kernels and inference tooling opened alongside it

This is the largest open weight model ever released.

[Lumina @LuminaXspace]

🚨Gemini 3.5 Pro is 100% getting A/B tested in the Gemini app

I have been trying to get a good example of two distinct differences in generation and got this A/B test.

The difference is night and day:

The right one is Gemini 3.6 flash, i got the same output in previous prompts trying to get an A/B window.

The left one is something else completely.

What do you think?

[Luke @luke_0302]

@LuminaXspace - Fable
- opus 5
- kimi k3 + Sol
- Grok 4.5
- Flash 3.6

[Lumina @LuminaXspace]

🚨Rank these AI models overall:

Only rank models you have actually tested yourself leave out the ones you have no idea about:

• Fable 5
• Opus 5
• GPT-5.6 Sol
• Kimi K3
• Grok 4.5
• Gemini 3.6 Flash
• Meta Muse Spark 1.1
• Opus 4.8
• GPT-5.6 Terra
• Sonnet 5

Drop your list in order based on your own experience.

No benchmarks or hype just which models have genuinely been best to use for you.

I’m collecting any results.

[Lumina @LuminaXspace]

🚨SpaceX has now signed the open weights AI letter

Microsoft’s updated signatory list now includes SpaceX alongside:

• NVIDIA
• Microsoft
• Meta
• Google
• OpenAI
• Hugging Face
• AMD
• Dozens of other AI and technology companies

Kimi K3’s full open weights are due today which lines up well too.

[Lumina @LuminaXspace]

🚨GPT-6 will drop soon:

Sam Altman is heading to Washington this week to preview OpenAI’s most powerful model yet👀:

• It's capable of original scientific discovery
• It solved an 80 year old maths problem autonomously
• Runs for much longer without constant supervision
• Powerful cyber capabilities (it's likely the model involved in the Hugging Face incident)
• This will have a new focus on “knowledge per dollar” rather than benchmarks alone

This model looks much bigger and better than a normal ChatGPT incremental update.

Are you ready for GPT-6?

[Lumina @LuminaXspace]

🚨July has been crazy for AI releases

Just look at what has dropped so far:

• Claude Opus 5
• Grok 4.5
• DeepSeek V4
• Kimi K3
• Qwen 3.8 preview
• GPT-5.6 Sol
• GPT-5.6 Terra
• GPT-5.6 Luna
• Gemini 3.6 flash
• Gemini 3.5 Flash-Lite
• Gemini 3.5 Flash Cyber
• Muse Spark 1.1
• Might have missed some too ?

And Kimi K3’s full weights are still due before the month ends.

The pacing of releases is getting ridiculously fast.

Do you think August will top this?

[Lumina @LuminaXspace]

🚨Google’s Gemma models have now passed 900 million downloads:

• This is an incredible milestone, especially as more major companies are starting to publicly defend open models.

• Gemma 4 alone has reportedly passed 300M downloads.

• This is a huge step in the right direction for AI in my opinion.

Have you ever tried a Gemma model?

[Lumina @LuminaXspace]

🚨Which AI subscription currently offers the best usage limits for $20–$30/month in your opinion?

Curious to hear why too?

[Lumina @LuminaXspace]

🚨 Claude Opus 5 could still launch today

Cursor leaks show a claude-opus-5-thinking-high string:

• First literal Opus 5 identifier seen in Cursor outside of the "Honeycomb EAP" leak
• Appeared inside a Max Mode error
• Safety fallback routing to Opus 4.8
• Anthropic still has not officially announced the model

Do you think Opus 5 drops today?

[Lumina @LuminaXspace]

🚨 Kimi K3 open weights are due in 3 days

Moonshot are releasing the full model files by July 27:

• 2.8T total parameters
• Native vision capabilities
• 1M token context window
• Built for coding, reasoning and long-running agents
• Technical details expected alongside the weights

This will become the biggest open weights AI release ever.

[Elon Musk @elonmusk]

Grok 4.5 is not quite as good as Fable, but it is very fast, cost-effective and gets the job done

[Lumina @LuminaXspace]

🚨 Google has revealed the first details about Gemini 4

Sundar Pichai says Google is building a significantly larger frontier model and could return to releasing new Gemini models almost every month.

What is confirmed:

• Gemini 3.5 Pro remains in partner testing
• Gemini 4 is described as a significantly larger and highly ambitious model
• They plan to attempt releases almost every month
• Google is prioritising coding and autonomous agents

Google is doing frequent Flash releases while preparing Gemini 4 as its real attempt to reclaim the frontier.

Do you think Gemini 4 can close the gap with OpenAI and Anthropic?

[Lumina @LuminaXspace]

🚨Gemini 3.6 Flash just entered the top four on OSWorld-Verified:

• Google’s new Flash model scores 83.0% on agentic computer use, up from 78.4% on Gemini 3.5 Flash.

• Google has also made computer use a built-in client-side tool through the Gemini API.

• Gemini 3.6 Flash doesn't dominate every intelligence benchmark, but its agentic computer use performance is absolutely incredible at its price.

I feel like it makes sense that Google models would be good as browser and desktop agents?

[Lumina @LuminaXspace]

🚨Gemini 3.6 Flash is now #2 on MLE-Bench scoring 63.9%