← All briefings
📰

subtl daily briefing

Share𝕏inEmail
AllAICryptoSaaSFinanceStartups

Good morning, tech enthusiasts. OpenAI is asking the government whether AI labs can collectively pump the brakes, DeepSeek just made long-context AI absurdly cheap, and your next sleep aid might literally smell like a pine forest. Here's everything you need to know.

In today's briefing

  • 1.OpenAI seeks antitrust clarity on AI slowdowns
  • 2.DeepSeek's memory-light model breakthrough
  • 3.AI-scented sleep tech sniffs out insomnia
  • ⚡Quick hits on other news
Latest Developments
News

🤖OpenAI to Congress: Can We All Just Slow Down AI Together?

The Rundown: OpenAI is asking whether AI labs can legally coordinate an industry-wide slowdown without running afoul of antitrust law.

The details:

  • ●OpenAI asked Congress whether an agreed AI development pause would be considered illegal collusion under antitrust rules.
  • ●Anthropic's latest threat report details real-world operations using AI for espionage, weapons guidance, and romance scams.
  • ●The question highlights a growing tension between AI safety coordination and competition law in the industry.
Why it matters: For founders, this is a litmus test for how the industry will handle future risks. If regulators bless coordinated safety pauses, we could see more joint governance—but ambiguous antitrust rules could also discourage legitimate safety collaborations. Keep an eye on how this settles.

📰 Source: The Neuron

Share𝕏inEmail
News

⚡DeepSeek's New Model Is Twice as Big, Yet 4x Cheaper in Memory

The Rundown: DeepSeek V4.1-Flash delivers a 552B-parameter MoE model that is nearly twice the size of its predecessor but uses 4x less KV cache per token.

The details:

  • ●The model uses just 890 bytes per token of KV cache, a 4x reduction versus V4-Flash and 1/8th the SSD cache storage.
  • ●New Compressed Sparse Attention 2 lets layers share KV and retrieval results, with a Hierarchical Sparse Indexer limiting search to a 2,048-block candidate pool.
  • ●It activates only 8B parameters per input token and 16B per output token via a Causal Encoder-Decoder architecture.
Why it matters: Efficient long-context serving is the key to scalable AI agents. DeepSeek's architectural shortcuts drop serving costs for long-context workloads to a quarter of competitors per task, giving startups a cheaper path to building agentic products—and forcing incumbents to match.

📰 Source: AlphaSignal

Share𝕏inEmail
News

😴Smell Ya Later, Insomnia: AI-Tuned Scents Hit the Bedroom

The Rundown: A wave of sleep tech startups is using AI-tuned scents delivered through the nose to help people fall and stay asleep.

The details:

  • ●Sleep tech startups are using AI to tailor scent blends in real time throughout the night.
  • ●Devices deliver aromas through the nose, tapping into the olfactory system to modulate sleep cycles.
  • ●The approach is part of a broader boom in consumer sleep technology, with startups combining scent diffusion with sleep tracking algorithms.
Why it matters: Founders in consumer health should note that personalization is moving beyond wearables into the senses. AI-driven scent delivery could open up new revenue streams, especially as sleep remains a top startup vertical.

📰 Source: The Hustle

Share𝕏inEmail
⚡

Everything else in the news today

Google DeepMind's fruit-fly connectome inspired projects like a Doom-playing fly and the FLM fly-language model.→
Florida's PriorityEvac will fly you out of a hurricane's path for $1.25k, selling 'peace of mind' ahead of storm season.→
After 71 years, Columbia House mails its final CD to the music club's remaining subscribers.→
The miniature housing niche is now a $2.5B market.→
'BTC is the purest AI trade' — the latest take from Milk Road on why Bitcoin might be an AI proxy.→
A U.S. bank settled an internal payment with its own token, signaling growing adoption of institutional stablecoins.→
SaaStr asks if the era of the sales-guy CEO is over, sparking fresh debate in founder circles.→
📰Briefing📈Intelligence📚Library