tampa criminal attorney

In comparison to last year=E2=80=99s Apple Intelligence splash, the comp= any led with a more classic, engineering-first, developer-forward and ecosy= stem-focused slate at WWDC yesterday.

This handful of smaller announcements, say IBM experts, will help set th= e stage for real-life, built-in AI. Think: better integration of AI in pers= onal contexts and across devices, as well as enhancements that are useful i= n everyday life, such as the ability to translate conversations.

=E2=80=9CThe thing Apple is trying to do in this keynote is subtly bring= all their platforms together and make generative AI just part of the gener= al workflow,=E2=80=9D says Chris Hay, a Distinguished Engineer at IBM, in a= n interview with IBM Think.

Apple=E2=80=99s new Foundation Models framework allows developers to tap into Apple=E2=80= =99s on-device intelligence. =E2=80=9CGiving developers access to Apple=E2= =80=99s AI models and the ability to play with and fine-tune these models i= s really powerful,=E2=80=9D says Kaoutar El Maghraoui, a Principal Research= Scientist at IBM, in an interview.

Apple=E2=80=99s approach is certainly steady and cautious, she adds. But= embedding AI in its ecosystem might be a winning strategy. =E2=80=9CThey a= re building their ecosystem right,=E2=80=9D Hay says. =E2=80=9CEverything i= s becoming joined together. They have the best hardware, and nothing beats = Apple=E2=80=99s silicon for on-device compute. All these features will just= get better and become part of the operating system.=E2=80=9D

=E2=80=94 Anabelle Nicoud

In only a few months, Anthropic=E2=80=99s Model Context Protocol (MCP) has gained major endorsements across the = tech industry, from Google=E2=80=99s Sundar Pichai to OpenAI=E2=80=99s Sam Altman, who announced earlier this spring that MCP would be suppor= ted across OpenAI products.

=E2=80=9CI believe 2025 is the year agents hit production,=E2=80=9D said= Sarmad Qadri, Co-Founder and CEO of agent evaluation platform LastMile AI, during an AI agent meet-up organized in San Francisco by = IBM and Neo4j. =E2=80=9CBecause of better models, test-time compute and rea= soning models, a lot of complexity is shifting left, into the inference lay= er.=E2=80=9D

Qadri added that in a way, =E2=80=9Csimple agent patterns is all you nee= d.=E2=80=9D And as agents scale, he noted, so will the need for MCP. As pro= of of MCP=E2=80=99s success, an MCP Registry Registry has already appeared. =E2=80=9CIt=E2=80=99s insa= ne,=E2=80=9D said Michael Hunger, VP of Product Innovation at Neo4j, during the event.=20

But enabling seamless agent communication remains a challenge=E2=80=94on= e that IBM=E2=80=99s Agent Communication Protocol (ACP) aims to solve. =E2=80=9CACP provide= s a unified interface for agents to communicate regardless of their framewo= rks and is openly governed, not just open source,=E2=80=9D explained Sandi = Besen, an AI Research Engineer and Ecosystem Lead at IBM, during the event.=

ACP was developed as part of the open-source BeeAI framework and is designed to standardize how agents communicate across systems=E2=80=94turning siloed agents in= to interoperable agentic systems. Google, too, has entered the space, launc= hing A2A, an agent-to-agent protocol that aims to promote agent collaborati= on and greater autonomy.

=E2=80=9CMCP and agent protocols like ACP and A2A are [complementary],= =E2=80=9D wrote Besen in a blog post. =E2=80=9CThere=E2=80=99s no universally = =E2=80=98best=E2=80=99 protocol=E2=80=94the right choice depends on your sp= ecific system architecture, environment, use case and goals.=E2=80=9D

=E2=80=94 Anabelle Nicoud

Apple researchers found that large reasoning models, such as Claude 3.7 Sonnet and DeepSeek-R1, drop to zero accuracy on more cha= llenging puzzles, revealing a sharp failure point in their ability to scale= logical reasoning, even with increased token counts and computational reso= urces.

FutureHouse unveiled ether0, a compact open-source reasoning mode= l trained on nearly 500,000 chemistry questions, which outperformed GPT-4.1 and DeepSeek-R1 on complex molecular tas= ks.

Google added interactive charts to AI Mode in Labs, letting = users visualize and compare stock and mutual fund data using Gemini= =E2=80=99s multimodal reasoning, turning complex finance queries i= nto dynamic, personalized graphs.

EleutherAI released the Common Pile v0.1, an 8TB dataset of publ= ic domain and openly licensed text designed to support open and transparent= training of large language models.

Was this email forwarded to you? You can subscribe here