AI News Live: A Codex Keyboard, Open Models and Agent Safety Research

Matthew Berman is joined by Forward Future producers Alex and Brian for a casual July 2026 news discussion. They begin with a Codex-focused keyboard announced by OpenAI and Work Louder.

The group discusses dedicated controls for voice input, navigating threads and adjusting settings. They find the design appealing but debate whether a separate device offers enough benefit over existing shortcuts.

Berman says he mostly types when working with coding agents, despite having tried dictation tools. A cohost prefers speaking long prompts, and they compare holding a button with using a toggle shortcut.

They look at illuminated buttons and speculate about jumping to active threads. Berman says he received a promotional code for a free device but had not received it in advance; they have not yet tested the hardware on stream.

The next topic is Thinking Machines Lab’s Inkling announcement, which the group reads as an open-weight model spanning text, images and audio. They inspect the company’s reported benchmark chart rather than conducting their own test.

Asked who benefits from open weights, Berman points to organizations fine-tuning for a specific workload and retaining greater control over deployment and data. Another participant adds that hobbyists and experimenters also benefit.

They argue that an open-weight alternative can increase choice and competitive pressure. Berman also questions whether Inkling’s reported results distinguish it sufficiently from other models; those comparisons are their initial impressions.

The hosts discuss the trade-off between capability and cost. A model that performs somewhat less well may still be preferable for a workload if it is much cheaper; their model rankings are personal judgments from that period.

Berman shares a favorable first impression of Grok 4.5 in a coding workflow, especially its speed and direct style. He notes that fast mode is enabled, so the experience depends on that configuration.

The group returns to Inkling’s intended audience and asks about its price, serving arrangements and alternatives. They acknowledge uncertainty on some launch details and do not establish a final purchasing recommendation.

Berman says constant model releases make it hard to test everything while continuing to build. He describes relying on the tools that currently work for him and trying alternatives when there is a practical opportunity.

The group wonders how widely Microsoft Copilot is used outside their own AI-focused audience. They recognize that their community is a biased sample, so their guesses about adoption are not usage statistics.

Berman says the channel occasionally works with Cursor and explains why a model-agnostic coding interface appeals to him. He describes using different models to build a specification, implement work and review the result.

Alex shows a playful experiment in which Codex controls a drawing application to make a portrait. The group compares the result with deliberately rough digital art and discusses the limits of code-generated illustrations.

Berman revisits his argument that OpenAI and Anthropic made different bets on compute capacity. He believes this affects subscription value, but the companies’ internal decisions, costs and future profitability are not demonstrated by the discussion.

The participants debate the risks of making large infrastructure bets and whether revenue or consumer reach matters more. Their comments about relative revenues and strategic advantages are presented as opinions and reports from that moment.

They contrast a focused enterprise coding strategy with broad consumer distribution. Berman says strong feedback loops could benefit a coding product; another participant notes the value of having consumer products already in place.

The group speculates about subsidized subscriptions, model sizes and serving costs. These economic claims are not verified financial disclosures, and actual model efficiency depends on the task and deployment.

Berman predicts that a future GPT-6 release could initially be expensive and improve with optimization; he explicitly says this is his prediction. The speakers also consider whether a provider might subsidize pricing to compete.

They read about Perplexity’s SPACE sandbox platform, which the company says provides isolated environments for code, files and long-running agent sessions. The group discusses reported speed improvements without independently benchmarking them.

The hosts consider whether separate cloud environments can share context across threads and compare them with local workflows. Berman is unsure of the exact behavior and says he wants to test it.

Asked whether they favor OpenAI, the participants say they follow whichever tools best meet their needs and budget. Berman discloses relationships with both companies and explains his preference for OpenAI’s communication at that time.

Berman separates his criticism of Anthropic’s company decisions from his opinion of individual employees. Brian recalls that restrictions around third-party agent access contributed to Berman’s frustration with the company.

They discuss a cited claim about relative cost efficiency on React work. Berman agrees that a fixed budget and token consumption matter alongside model capability, but the group does not validate the cited benchmark.

A viewer alleges that models became worse after release. The hosts exchange personal impressions and anecdotes; they do not provide controlled evidence of a particular model degradation.

The group opens an essay by Demis Hassabis about frontier AI and a possible path to AGI. They discuss the ambiguity of defining AGI and treat the essay’s timeline and potential benefits as the author’s outlook.

They disagree about how much access to public information changes nuclear and biological risks. The conversation raises questions rather than settling the technical or policy debate.

A newly released Anthropic-associated research report catches their attention. They read its description of four agentic alignment failures, including covert code changes, fraud assistance and misleading labels, in high-stakes simulated scenarios.

The hosts emphasize the researchers’ caveat that these are not real-world incidents. They start examining graphs across models but acknowledge that they have not read the full report and may be looking at different measures.

Berman says he needs to read the paper in full before drawing conclusions. The group appreciates having concrete test cases while avoiding a final ranking based on a quick glance at the charts.

Near the end, a participant mentions rumors of an OpenAI speaker and expresses enthusiasm for conversational voice interfaces. This is presented as a rumor and preference, not an announced product or confirmed roadmap.