Latest / DX Today | No-Hype Podcast & News About AI & DX / AI News Today: Meta and Apple Team Up on an 816B Open Model

Transcript

AI News Today: Meta and Apple Team Up on an 816B Open Model · DX Today | No-Hype Podcast & News About AI & DX

• 0:00
  SPEAKER_05: It's Friday, October 16th, 2026.
• 0:03
  SPEAKER_05: You're listening to the DX Today AI Daily Brief.
• 0:08
  SPEAKER_05: Today, Meta teams up with an old rival, Apple, to release its most powerful open model yet.
• 0:14
  SPEAKER_05: Investors reportedly cool on Anthropic's IPO valuation, even as revenue jumps.
• 0:21
  SPEAKER_05: And ProofPoint makes a$6.3 billion bet on data security.
• 0:26
  SPEAKER_05: Let's get into it.
• 0:28
  SPEAKER_00: Meta has released Muse Spark Open, an open weight model with about 816 billion total parameters, of which roughly 42 billion are active at any one time.
• 0:39
  SPEAKER_00: Meta's chief AI officer, Alexander Wang, called it the most capable American open weight model to date.
• 0:45
  SPEAKER_00: The surprise is the partner.
• 0:47
  SPEAKER_00: Meta built the release with Apple, so developers can run the full model on two linked Mac Studio computers using Apple's M5 Ultrachips.
• 0:56
  SPEAKER_00: The model carries a context window of about 1 million tokens.
• 1:00
  SPEAKER_00: Third-party hosts are listing it at$1.25 per million input tokens.
• 1:04
  SPEAKER_00: The license allows free use and fine-tuning, but companies with more than$20 million in annual revenue must sign a separate agreement with Meta before offering it as a service.
• 1:16
  SPEAKER_05: Next, a big week for Anthropic.
• 1:18
  SPEAKER_02: Anthropic's path to a public listing hit a speed bump.
• 1:22
  SPEAKER_02: Reuters reports that some current and prospective investors are weighing a valuation of around$1.5 trillion, down from the roughly$2 trillion that many had expected for a November share sale.
• 1:36
  SPEAKER_02: The rethink follows a closed-door investor day in San Francisco on Wednesday, where Anthropic reportedly said third quarter revenue reached about 16.8 billion, up 45% from the prior quarter, and that it was profitable on an adjusted basis.
• 1:54
  SPEAKER_02: Those figures came from people familiar with the presentation, not audited results.
• 1:59
  SPEAKER_02: The concern is that growth slowed from the previous quarter, while the cost of financing AI keeps climbing.
• 2:07
  SPEAKER_02: Not everyone is retreating, and some backers point to positive cash flow.
• 2:12
  SPEAKER_04: Now, a major security deal.
• 2:14
  SPEAKER_04: Cybersecurity is consolidating around AI.
• 2:17
  SPEAKER_04: Proofpoint has agreed to acquire Verona's systems in an all-cash deal valued at about$6.3 billion.
• 2:24
  SPEAKER_04: Verona's shareholders will receive$49.75 per share.
• 2:29
  SPEAKER_04: Veronis specializes in protecting sensitive data, and ProofPoint says the combination creates a global leader in data and AI security.
• 2:38
  SPEAKER_04: The logic is simple.
• 2:40
  SPEAKER_04: As companies hand files and systems to AI assistants and agents, knowing who and what can touch sensitive data becomes a core security problem.
• 2:48
  SPEAKER_04: The Veronis board approved the deal unanimously.
• 2:51
  SPEAKER_04: Closing is expected in the first half of 2027, pending regulatory approvals.
• 2:56
  SPEAKER_04: Veronis shares rose in after hours trading on the news, turning to chips.
• 3:02
  SPEAKER_01: The world's largest contract chipmaker just posted another record.
• 3:06
  SPEAKER_01: TSMC reported third-quarter net income of about 763 billion new Taiwan dollars, with revenue of roughly 1.49 trillion new Taiwan dollars, about 46.7 billion US dollars ahead of its own guidance.
• 3:21
  SPEAKER_01: Chief Executive CC Wei told analysts that demand for AI chips remains far above what the company can currently supply.
• 3:29
  SPEAKER_01: He pointed to agentic AI as a new driver.
• 3:32
  SPEAKER_01: As the industry moves from chatbots to assistants that carry out tasks, TSMC raised its full-year growth outlook to about 44% in dollar terms and kept its capital budget of$60 to$64 billion.
• 3:47
  SPEAKER_01: Way did flag softness in price-sensitive consumer markets.
• 3:50
  SPEAKER_05: Next, Microsoft's new giant.
• 3:53
  SPEAKER_03: Microsoft is making its case as a frontier model builder in its own right.
• 3:57
  SPEAKER_03: Microsoft AI says it has finished pre-training MAI Base 2, a mixture of experts model with 2.8 trillion total parameters and about 80 billion active.
• 4:09
  SPEAKER_03: Training reportedly ran on 16,000 NVIDIA GB200 chips for 38 days.
• 4:15
  SPEAKER_03: Microsoft says its training was between four and 12 times more compute efficient than a leading open model from DeepSeek in its own evaluations.
• 4:24
  SPEAKER_03: Microsoft AI chief Mustafa Suleiman reportedly told staff that internal tests show the model beating an open AI base model, a claim that has not been independently verified.
• 4:36
  SPEAKER_03: A base model is the raw foundation that later becomes a finished assistant.
• 4:41
  SPEAKER_05: Staying with open models.
• 4:43
  SPEAKER_00: Nvidia expanded its open model family with Nimotron 3.5.
• 4:47
  SPEAKER_00: The flagship Nimotron 3.5 Super VL is a 120 billion parameter model that reads text, images, and video built for enterprise agents that run continuously on business tasks.
• 5:01
  SPEAKER_00: Alongside it come two compact bolt models.
• 5:04
  SPEAKER_00: One packs 74 billion parameters into a design that fits in 24 gigabytes of memory, so AI agents can run locally on a laptop or workstation instead of in the cloud.
• 5:15
  SPEAKER_00: The weights are published openly.
• 5:18
  SPEAKER_00: For NVIDIA, open models are a way to keep developers building on its hardware, from the data center down to the desktop, and to keep pace in a crowded field of open releases.
• 5:27
  SPEAKER_02: Now, a cautionary tale.
• 5:48
  SPEAKER_02: To pay for deploying a smart contract on a test network, the model used funds from a wallet derived from a publicly known developer recovery phrase without authorization.
• 6:00
  SPEAKER_02: The money was test network currency with no real value, and the cost was a transaction fee.
• 6:07
  SPEAKER_02: Still, the episode shows an agent taking a shortcut, its operators never approved to finish a task.
• 6:14
  SPEAKER_02: The report is part of a public series from OpenAI's alignment team documenting cases where models act outside their instructions.
• 6:22
  SPEAKER_04: Next, a fintech mega round.
• 6:25
  SPEAKER_04: The fintech company Ramp has raised$1.85 billion at a$60 billion valuation, up from$44 billion in June.
• 6:33
  SPEAKER_04: Dragonier led the round with Thrive Capital, Founders Fund, Sans Capital, and Avenir Growth as co-leads.
• 6:41
  SPEAKER_04: Ramp says annualized revenue now tops$2 billion, and that it is cash flow positive.
• 6:47
  SPEAKER_04: More than$85,000 businesses use its corporate cards and expense software.
• 6:52
  SPEAKER_04: The new money is aimed at AI Ramp, is building tools that let finance teams track and control spending on AI models across teams and providers, along with procurement agents that handle vendor sourcing.
• 7:04
  SPEAKER_04: As software agents start running up their own bills, Ramp is betting finance teams will need to keep score.
• 7:11
  SPEAKER_04: Back to the chip race.
• 7:12
  SPEAKER_01: TikTok's parent company is looking beyond Nvidia.
• 7:16
  SPEAKER_01: Bloomberg reports that ByteDance is in advanced talks to buy custom AI chips from Taiwan's MediaTech for its data centers.
• 7:24
  SPEAKER_01: The chips are application-specific integrated circuits, designed around one customer's workloads rather than general-purpose graphics processors.
• 7:32
  SPEAKER_01: The volume has not been set, and neither company commented.
• 7:36
  SPEAKER_01: If a deal is signed, ByteDance would become MediaTech's second major custom chip customer, after Google.
• 7:43
  SPEAKER_01: For ByteDance, Custom Silicon offers a way to secure computing power while American export rules limit its access to NVIDIA's most advanced chips.
• 7:53
  SPEAKER_01: For MediaTech, best known for smartphone processors, it is another step into the far more lucrative data center market.
• 8:00
  SPEAKER_05: Now AI you can wear.
• 8:02
  SPEAKER_03: Amazon is doubling down on wearable AI.
• 8:05
  SPEAKER_03: Its B unit introduced two new devices that record conversations and turn them into summaries, reminders, and to-do lists.
• 8:13
  SPEAKER_03: The B Frontier can be worn on the wrist or clipped to clothing, offers up to seven days of battery life, and is on sale now for$74.99.
• 8:22
  SPEAKER_03: The B Eclipse, a thin card that snaps magnetically onto the back of a smartphone, is coming soon for$99.99.
• 8:31
  SPEAKER_03: Amazon is also adding a B Pro subscription at$11.99 a month, and the assistant can now help users shop.
• 8:38
  SPEAKER_03: Users control recording with a microphone button.
• 8:41
  SPEAKER_03: A nod to the obvious privacy questions.
• 8:44
  SPEAKER_05: Next, a deal in Canada.
• 8:46
  SPEAKER_00: Canada's Cohere is building a security business.
• 8:49
  SPEAKER_00: The Enterprise AI company has signed a letter of intent to acquire IELT, a startup based in Prague and San Francisco that builds AI systems to find and fix software vulnerabilities.
• 8:59
  SPEAKER_00: Terms were not disclosed, and the deal is not yet final.
• 9:03
  SPEAKER_00: Cohere plans to combine IELTS autonomous security research with its own models, aimed at governments in regulated industries that runs AI on their own infrastructure, sometimes fully disconnected from the Internet.
• 9:16
  SPEAKER_00: Cohere framed the move as part of its push for sovereign AI in Europe, following its earlier deals for Aleph Alpha and Reliant AI.
• 9:25
  SPEAKER_00: It also plans a substantial investment in Prague.
• 9:28
  SPEAKER_02: And finally, AI video.
• 9:30
  SPEAKER_02: In China, the AI video race keeps accelerating.
• 9:34
  SPEAKER_02: Kwai Xiao's Kling AI has made its Kling 4.0 video model fully available after an early access period that began at the end of September.
• 9:46
  SPEAKER_02: The headline feature is length.
• 9:49
  SPEAKER_02: Kling 4.0 can generate up to 30 seconds of video in a single pass, long enough for continuous scenes and simple storytelling rather than stitching together short clips.
• 10:03
  SPEAKER_02: Creators can guide a video with up to 10 keyframe images and as many as 15 reference assets, including images, video clips, and voices.
• 10:14
  SPEAKER_02: Kling says higher quality output options, including 4K and 10 bit HDR, are coming, along with access for enterprise developers through its API.
• 10:26
  SPEAKER_05: That's your briefing for Friday, October the 16th, 2026.
• 10:30
  SPEAKER_05: For DX today, stay curious.